Benchmarking of Supervised Machine Learning Algorithms in the Early Failure Prediction of a Water Pumping System

Preventing failures in water supply systems is of vital importance for the development of a population, especially when its economic engine is the agricultural sector. Therefore, it is important to apply new control techniques, which incorporate machine learning and allow prediction failures effectively. This paper performs a comparative analysis of three classification algorithms, random forest, support vector machines, and artificial neural networks, to predict failures in a water pumping system. The methodology employed considers the selection of a training dataset, data preprocessing, training, and evaluation of each model, and its subsequent performance comparison. According to the results, the lowest average accuracy was obtained by the SVM algorithm (83.24%), while RF obtained the highest accuracy (99.98%), closely followed by ANN (86.94%). According to the hypothesis tests, there are significant differences between the SVM, RF, and ANN algorithms, showing that the latter two achieve better performances than SVM, but without significant differences between them, so that to select one of them, it is necessary to consider other aspects such as training time and interpretability. The results show that supervised learning algorithms can reach values higher than 80% of accuracy in the detection of system failures, which evidences their usefulness in control systems.