Zhou, S; Coleman, D; Turnbull, A (2001) Development of neural networks to predict general corrosion of duplex stainless steels. NPL Report. MATC(A)22

Text MATC22.pdf Download (245kB)

Abstract

Multi-layer feed-forward (MLF) neural networks have been applied to data from the NPL Corrosion Database of Duplex Stainless Steels in order to predict the corrosion rate of DSS in sulphuric acid, formic acid and hydrochloric acid. Although the amount of data available looks at first glance to be quite significant, on more detailed examination, it is clear that the amount of data in the conditions of importance, i.e. at the boundaries between low, medium and high corrosion rates, is actually quite sparse. Nonethe less, the use of Kohonen neural networks to 'focus in' on the important data has allowed the development of predictive algorithms although the detailed validation of these is not possible, and therefore care should be taken when using them.
Analysis of the MLF neural networks has been carried out based on the network weights. These indicate the relative importance of the input parameters viz, pitting resistance equivalent number (PREn),concentration of aggressive environment and temperature. Neural-fuzzy algorithms that "mimic" the MLF neural networks have also been developed. Although the performance of these networks (in terms of distinguishing the classes of data, viz. low medium and high corrosion rates) is not as good as the MLF neural networks, they have produced some "rules" which do describe the basic behaviour of the duplex stainless steels.

Item Type:	Report/Guide (NPL Report)
NPL Report No.:	MATC(A)22
Subjects:	Advanced Materials Advanced Materials > Corrosion
Last Modified:	02 Feb 2018 13:17
URI:	http://eprintspublications.npl.co.uk/id/eprint/1950

Actions (login required)

View Item