H.M.G Smets

Deriving corrosion knowledge from case histories: the neural network approach

Abstract In the corrosion domain, a number of case histories describing different failure situations of several materials under a variety of environmental conditions exist. Although, perhaps every case history in itself does not provide so much valuable information, knowledge is certainly present when considering all the case histories. This source of information is however, virtually untouched, mainly because suitable technologies for the processing of this kind of information have only recently evolved. In this paper, it will be shown how valuable information can be extracted from case histories by means of neural networks that are capable of extracting regularities from sets of examples which can then be used for the processing of previously unseen instances. Based on case histories of service experience with austenitic stainless steels in chloride-bearing water, two neural networks have been developed so far to predict the chloride-induced stress corrosion risk in an intelligent, automated way. The first one reflects the temperature and chloride concentration dependency whereas the second one considers the influence of the oxygen concentration and chloride concentration in high temperature water.

Classification of Corrosion Images by Wavelet Signatures and LVQ Networks

In this paper, a method is described for the classification of corrosion images into two distinct classes. Since segmentation is very difficult, an automatic feature selection and classification procedure is preferred. This is done by performing a wavelet decomposition of the images, and computing energy signatures from the decomposition. Compact signature vectors represent the images and effectively characterize their type. The recognition is performed with a Learning Vector Quantization network. The method is tested on a set of 398 images, 260 of which were for training. High recognition scores are obtained.

SCC susceptibility analysis of stainless steels in nuclear reactor water: a neural network and expert system approach

Abstract Stress corrosion cracking is one of the most troublesome phenomena encountered in boiling water nuclear reactors. The sensitivity of stainless steels to stress corrosion cracking in high-temperature water can be modelled by means of neural networks. The neural network described in this paper learned to recognise the combined effect of temperature, chloride and oxygen concentration on the occurrence of stress corrosion cracking. In order to yield a more thorough corrosion risk assessment also taking into account additional factors, such as stress level, degree of sensitisation, etc., the latest available expert system technology, which offers the possibility of accessing different knowledge sources (rules, mathematical models, neural networks, ...) at the same time, is put to use.