Improving Pipe Degradation Modeling Using Artificial Intelligence Models with Age Plus Other Variables

Abstract

This paper uses data on stormwater pipes from 21 distinct ages ranging from 3 to 110 years, the degradation of which had been originally modeled using a Markov scheme (using only age as an explanatory variable); in order to establish whether artificial intelligence (AI) models can improve the fitting and prediction of degradation levels. The AI models comprise (1) an artificial neural network (ANN) with only age as a variable; (2) another ANN model with pipe material, pipe diameter, exposure condition, and soil type as variables in addition to age; and (3) a random forest (RF) model with the same five variables. The multiparameter models, especially the RF one, were found to perform much better (e.g., R2 value of 0.98) than those using only age as the independent variable (with R2 values around 0.24). The RF model also showed better performance with respect to time stationarity. Pipe material was found to be an explanatory variable almost as dominant as age, with concrete pipes deteriorating much less than vitreous clay ones. The fluctuation of the observed overall damage rating with age was reflected in a similar pattern of fluctuation for the proportion of clay pipes in the samples obtained at the different ages.