The Effects of Socioeconomic Status, Geological, Engineering, and Climate Factors on Machine Learning-Dependent Forecast of Water Pipe Failure

Abstract

Subterranean water pipes deteriorate due to a variety of physical, mechanical, environmental, and social factors. Accurate pipe failure prediction is a prerequisite for a rational administrative approach to the water supply network (WSN) and is a challenge for the conventional physics-dependent model. The enormous water supply network's past maintenance data history was utilized by the study to anticipate water pipe breaks using data-directed machine learning methodologies. In order to include many factors that contribute to the deterioration of subsurface pipes, an initial multi-source data-aggregation system was created. The framework outlined the conditions for merging many datasets, including the soil type, population count, geographic, and meteorological datasets as well as the conventional pipe leaking dataset. Based on the data, five machine learning (ML) techniques—LightGBM, ANN, logistic regression, K-NN, and SVM algorithm—are developed to forecast pipe failure. It was found that LightGBM provided the optimum performance. Five criteria were used to analyze the relative importance of the main contributing factors to the water pipe breakdowns: calculation time, accuracy, effect of categorical variables, and interpretation. LightGBM, the model with the second-lowest training time, performed the best. Given the severe skewness of the dataset, it has been shown that the receiver operating characteristics (ROC) measure is too optimistic using the precision-recall curve (PRC) metric. It's noteworthy to note that socioeconomic factors within a community have been shown to have an impact on pipe failure probability. This study implies that data-directed analysis, which combines ML techniques with the proposed data aggregation architecture, may enhance reliable decision-making in WSN administration.