Development of a Neural Network Model for Predicting Damage Functions for Pitting Corrosion in Condensing Heat Exchangers

Pitting attack occurs in ca. 10% of domestic and industrial gas fired heat exchangers, and generally appears during the first five years of operation. The causes of pitting corrosion are several, including the use of chlorinated solvents in the ambient environment, the quality of the gas burned, and the material used to fabricate the heat exchanger. Several attempts have been made to develop predictive models based upon observed pitting data, but they are limited in their predictive capabilities. Recently, we have initiated a program to develop a deterministic model to predict the damage resulting from pitting corrosion. However, the problem is complicated, and several restrictive assumptions have had to be made to render the problem tractable. An alternative approach, which is developed here, is to assume that we have no intrinsic information concerning the physico-chemical mechanisms involved in the nucleation and growth of pits, but that we are able to discern relationships between the observed damage and various input parameters which may be used to extrapolate the damage to future times. Probably the most efficient method of establishing these relationships is to use artificial intelligence techniques. Accordingly, we describe here an Artificial Neural Network (ANN) for predicting pitting damage functions for condensing heat exchangers. When the net is trained with reliable data and knowledge, we are able to predict accurately damage functions under significantly different conditions.