Investigating the Impact of Cumulative Pressure-Induced Stress on Machine Learning Models for Pipe Breaks

Significant financial resources are needed for the maintenance and rehabilitation of water supply networks (WSNs) to prevent pipe breaks. The causes and mechanisms for pipe breaks vary between different WSNs. However, it is commonly acknowledged that the operational management and water pressure influence significantly the frequency of pipe breaks. Pipe breaks occur when the water pressure exceeds the tensile strength of a pipe, or due to repetitive pressure cycles that result in fatigue-related failures. Considering these pipe failure modes, a new metric known as cumulative pressure-induced stress has been introduced. This metric takes into account both static and dynamic pressure components that contribute to pipe breaks, including mean pressure and the magnitude and frequency of pressure fluctuations, respectively. The impact of CPIS on pipe breaks has not been extensively investigated. Consequently, this study investigates and evaluates the impact of this metric when incorporated as an explanatory variable in Random Forest (RF) models that analyse the key causes of pipe breaks in two WSNs. Different RF models were developed both with and without incorporating pressure components. Subsequently, the performance of these models and the significance of each input variable were assessed. The results of this study suggest that CPIS is an important variable, especially in cases where pressure-related factors play a significant role in pipe breaks. Consequently, incorporating CPIS has shown a notable improvement in the accuracy of pipe break models.