Abstract
This paper identifies axial stress decrease of six bolts attached to a rotational body. For each bolt head a distance from a sensor probe is measured while the rotational body is under rotation. A characteristic parameter representing a bolt head and a flange deformation is calculated for each bolt from the measured distance. Three types of diagnosis problems are dealt with by a multi-layer neural networks approach. A set of characteristic parameters and the number of hidden layer's neurons are changed and their effects on diagnosis performance are investigated. It turns out that the total of six bolt head deformation parameters are useful for the diagnosis, yielding low failed-safe and failed-dangerous probabilities provided that the number of hidden layer's neurons are suitably determined. The neural network gives better diagnosis performance than Bayes' discriminant function approach.
