Performance comparison for pipe failure prediction using artificial neural networks

Infrastructure managers must decide on the replacement timing of buried pipes in water distribution networks. These assets deteriorate resulting in failures and their associated consequences. In recent years, studies have investigated failure prediction models for pipes based on artificial neural networks (ANNs). These models are either generalized (i.e. trained with all pipe failures) or specialized (i.e. trained with certain pipe failures based on pipe material or failure history). It is currently unclear whether prediction accuracy is improved by developing several specialized ANNs compared to a single generalized one. To answer this question, four modelswere developed: a generalized model for cast iron (CI) and ductile iron (DI) pipes; a specialized model for CI pipes, a specialized model for CI pipes with no previous failures and a specialized model for CI pipes, which had previously failed. Overall, the study found minimal difference in performance between the generalized and specialized models.