Leak detection and localization in pipeline networks using machine learning and principal component analysis

Tesi en modalitat de cotutela: Universitat Politècnica de Catalunya i Instituto Tecnológico de Tuxtla Gutiérrez (English) This thesis presents a methodology to locate leaks in pipeline networks, with a focus on water distribution networks. Leak localization techniques using classifiers are proposed by associating the different classes with branches or nodes where the leaks occur. The classifiers are fed with pressure measurements at specific sensor nodes in the network and learn to recognize the leak positions through supervised learning from a data-set obtained by the network model simulations. The analysis of the results includes tests in different leak scenarios and considers measurement noise in the pressures. The proposed classification techniques’ performance is analyzed in a didactic network and with physical measurements in an experimental prototype’s real networks. Performance metrics, classification loss (or its counterpart, accuracy), and topological error (measuring the distance between the leak’s real position and the position where it is detected) are used to validate the methods’ performance and applicability. Unlike other works as reported in the literature, this study improves the classifiers’ performance when a non-linear transformation is applied on the pressure residuals to extract only the information about their direction so that the classifiers use “direction cosines” as features. However, it was also found that performance using direction cosines can be considerably affected by uncalibrated models or sensors. This research also addresses a fundamental problem in leak localization-related to find-ing an optimal placement for the pressure sensors. This is a significant finding because the number of sensors to place is limited in real networks due to technological and economic problems. Two algorithms are proposed to select the nodes where it is most useful to place the sensors to capture as much information as possible about the location of the leaks within the network. The first is a supervised approach based on information theory that seeks maximum relevance and minimum redundancy of the sensor set in terms of mutual information. The second is an unsupervised method based on principal component analysis that aims to maximize the pressure variance associated with leaks captured by the sensors. Both are iterative algorithms independent of the leak localization method and are econom-ical in computational time compared to other proposal