Keywords neural networks, image processing, robotics, transport scheduling
Start Date: 15-NOV-88 / Duration: 36 months
[ contact / participants ]
The aim of ANNIE was to find out which of several generic problem areas, such as pattern recognition, sensor fusion, and adaptive control and optimisation, are best approached using neural networks.
To do this, the partners prototyped neural network solutions in simulated problems chosen from automatic inspection, condition monitoring, robotics, control and scheduling.
A survey on the systematics and capabilities of neural networks, their architecture and applications has been carried out, and a taxonomy published.
A survey of the software and hardware available for neural network simulation has also been performed. Areas of applications where neural networks might complete or out-perform conventional techniques have been selected:
Image processing and inspection
The inspection work is now largely completed, with a number of successes achieved. Two problem areas related to non-destructive testing have been chosen: visual inspection of solder joints, and ultrasonic testing of pressure vessels. In particular the SIEMENS solder joint work has demonstrated excellent performance. It is likely that more than one partner will actively pursue the technology as part of commercial products.
ANNIE's work stimulated a large internal programme in AEA on graph-matching, and now a package called "MatchFinder" has been produced which enables automatic matching and comparison of gas-liquid chromatography (GLC). The product is being marketed jointly with Perkin Elmer.
Neural networks can be of considerable help in the visualisation and analysis of multivariate data, as in the ANNIE application of the Harwell NetVision software. This was used in the analysis of both the nondestructive testing and Lehrfahrzeug data. The evaluation experience obtained in this work has helped to further the development of NETVision (renamed RENDER) as a tool for the pharmaceutical industry to help analyse structure-activity relationship in drug design. This product is now being tested by selected industrial users. Further applications to other industries are expected to follow.
The successful neural network methods developed within ANNIE for welding defect classification using ultrasonic diffraction images have direct application in the AEA's non-destructive testing activities, for nuclear plant, for offshore structures, and for other industrial applications where safety and plant integrity command a high premium. The AEA's National NDT Centre already uses and markets a sophisticated PC-based ultrasonic data acquisition and analysis system, which can accommodate neural network classification routines derived from the ANNIE software. In the manufacturing area, a prototype acoustically-based inspection technique for a tile production line has been developed jointly with a major UK tile producer.
A prototype of an autonomous roving vehicle was demonstrated at the 1989 ESPRIT Conference. This work has been found very useful in approaching the problem of collision avoidance.
Neural networks have also been found to be an ideal means of monitoring the oscillations of flexible structures such as vehicles, aircraft and spacecraft) and controlling them using a damping system based around a neural network.
An application for airline end-user and crew-scheduling problems has been investigated. A first prototype has been produced, and the follow-up work planned aims to produce a crew-scheduling product. Applications in banking and finance are also planned to exploit the project's results.
The handbook and the neurocomputers and hardware book should be published soon in the Springer Verlag ESPRIT series. A tutorial with around 60 participants was held during the 1991 ESPRIT Conference Week.
The partners have common interests in the application domains. They have the resources required to follow through pre-competitive work into the production systems of either a dedicated analogue device or an application-specific integrated circuit (ASIC). They are also in a position to contribute towards the further development of fully adaptive network hardware using optics and new materials technology.
Dr J.C. Collingwood
Material Physics/Metallurgy Division
UK - ABINGDON OX11 0RA
tel: + 44/ 235-24141 ext 5222
fax: + 44/ 235-235-432726
telex: 83135 ATOMHA G
AEA TECHNOLOGY - UK - C
ALPHA SAI - GR - P
BRITISH AEROSPACE PLC - UK - P
CETIM-ETABLISSEMENT DE SENLIS - F - P
IBP PIETZSCH GMBH - D - P
SIEMENS AG - D - P
KPMG PEAT MARWICK CONSULTANTS - D - A
ANNIE - 2092, December 1993
please address enquiries to the ESPRIT Information Desk
html version of synopsis by Nick Cook