Real-time detection of urban gas pipeline leakage based on machine learning of IoT time-series data

Abstract

China is advancing urban gas pipeline safety through the strategic deployment of methane detectors. This study introduces a systematic approach for identifying underground gas leaks, and overcoming challenges such as biogas interference and environmental factors. Abnormal methane sequences, defined by expert experience, undergo data preprocessing, including cleansing, compressing, and extracting rising segments. Feature engineering is performed on these rising segments across three dimensions: temporal, volatility, and temperature features. Different classification algorithms, including K-nearest neighbor, decision tree, and support vector machine, are employed for gas leakage recognition. The decision tree proves most effective, achieving a 92.86% recall rate for methane leakage segments. Additionally, feature importance ranking from the decision tree model highlights the maximum value of preprocessed methane concentration time series and the fitting slope of the rising segment as crucial features for methane leakage detection.