Pipe Feature Classification Using Support Vector Machine Based on Time-Domain Feature Extraction

Abstract

Water infrastructure deterioration threatens global water security, with utilities losing over $14 billion annually due to non-revenue water (NRW). Malaysia reported 37.1% NRW in 2023, significantly exceeding international benchmarks. This study develops an automated pipeline monitoring system integrating pressure transient analysis with machine learning for leak detection in water distribution networks. Three feature selection methods (Minimum Classification Error, minimum Redundancy Maximum Relevance, and Sequential Forward Selection) were systematically evaluated to optimize classification performance. The proposed MCE-SVM model, which is based on two time-domain features (Slope Sign Change (SSC) and Integrated EMG (IEMG)) to characterize three pipeline conditions, including elbow, junction, and leak, attained a test accuracy of 87.4%, substantially higher than those achieved by mRMR-SVM (77.3%) and SFS-SVM (65.8%), demonstrating the effectiveness of the MCE based method in selecting highly discriminatory features with a minimal feature set. Moreover, the suggested system is capable of fast processing of the signal, which highlights the appropriateness of the proposed system to the report on near real-time monitoring of pipelines. The controlled laboratory experiments were conducted using 300 transient pressure signals recorded with one configuration under 4 bar in one pipe setup (63 mm MDPE).