SIENA

Stimulation Initiative for European Neural Applications

Esprit Project 9811

Case Studies of Successful Applications

Diagnosis of Spot Welds.

Company Background

TNO is a Dutch, government-sponsored research organisation to support national industry in all aspects of innovation, design and production. TNO-Industry is a subsidiary that focusses on industrial production issues.

The problem

Spot welding is a popular technique for robotic construction in the mechanical industry. It is based on the local heating/cooling of the melt between two adjoining plates of metal. As the melt is invisible, the quality of welding can not be visually inspected with non-destructive methods. Then again destructive diagnosis can only be applied to random production samples. Hence for real monitoring of the production quality, the need for non-destructive methods remains. One of the potential techniques is based on ultrasonic reflections. This is currently used for off-line visual inspection and gives the experienced operator some indication which still must be augmented by the occasional destructive sample.

Neural network application

Measurement data were collected by using a hand-held ultrasonic signal source. It was established that the features used in current practice do not provide a training quality that exceeds that of the human operator. Additional neural screening was required to establish features on which the network could be trained for a better diagnosis quality. Such features have been identified; then a neural feedforward network has been trained by error backpropagation and validated in the InterAct environment to approximate the weld reflection characteristics. The final network is currently in discussion for introduction in automotive production environments.

Benefit

For mechanical constructs, the "zero-defect" strategy is of upmost importance. Production quality can only be cost-effective where destructive tests are ruled out. In-line non-destructive tests for assembly techniques that defy visual inspection can be enabled by knowledge-intensive methods such as neural networks. The search for features simultaneous with the learning of a feature-based model allows for a fast turn-around evaluation of a proposed technique.

Generalization

Many technical problems can be handled by expert rules. However, expert knowledge is often based on "common sense" in combination with "an easy work-around". The neural network technique provides a structured approach to the engineering of the solution space in combination of the rapid prototyping of the modelled solution. The diagnosis of spot welds is just one of the many stages in a production process where quality monitoring can be achieved by such methods.

Contact Person

Dr.Ir. J.A.G. Nijhuis, Rijksuniversiteit Groningen, Dept. of Computing Science, P.O. Box 800, NL-9700 AV Groningen (The Netherlands)

Reference for further reading

(terBrugge, M.H., Jansen, W.J., Nijhuis, J.A.G. and Spaanenburg, L.) Neural feature classification for ultrasonic weld testing, Proceedings of the GRONICS'95 (Groningen, The Netherlands, 24 Februari 1995) pp.7-10.