Machine Learning Based Water Pipeline Leakage Detection And Analysis

Abstract

Fluid transport systems are negatively impacted by pipeline leaks, which also provide serious problems to public safety and operational dependability. Conventional leak detection techniques are sensitive to interference, lack of scalability, and inability to reliably differentiate between various forms of leaks, relying on fixed thresholds, human inspection, or inferential logic.An intelligent, comprehensible, and data-driven framework for leak detection and categorization is developed in this paper combining unsupervised clustering, supervised learning, statistical feature engineering, and digital signal processing. This paper uses a database containing 1000 cases collected using three types of sensors (accelerometer, hydrophone, and pressure). Two from each sensor were used. Thirteen statistically significant physical descriptors are extracted by normalizing pressure and vibration signals that are collected from sensors. Binary leak detection is carried out in the first step of a hierarchical Machine Learning (ML) architecture, where a binary representation of (1) = Leak (all categories) and (0) = No Leak (NL) , then leaks identification are categorized into four groups: Circular Crack (CC), Longitudinal Crack (GL), Localized Leak (LC), and Orifice Leak (OL). Using structured pipeline datasets, Random Forest (RF) classifier supported by the K-Means clustering algorithm implemented and assessed. With an area under the Receiver Operating Characteristic (ROC) curve of 0.777%, this indicates that the model performs far better than haphazard guesswork and has a high degree of reliability in differentiating between normal and leak circumstances and distinct feature separation, the results showed successfully a practical leak detection system that achieving 94.64% accuracy through intelligent confidence thresholding.. A remarkable link between the learned patterns and actual physical leakage phenomena is confirmed by feature importance analyzing, correlation, confidence behavior, and threshold sensitivity supported with human-intervention-based decision-making techniques.