Development of a Predictive Maintenance Framework For Orifice Gas Metering Systems Using Machine Learning

Abstract

Accurate measurement of natural gas flow through orifice meters is essential for gas metering, custody transfer, and financial transparency; however, orifice meters are susceptible to faults such as erosion and particulate accumulation, which can lead to inaccurate readings. This research developed a practical, physics-informed unsupervised learning framework for early problem detection using real field data collected from January through September 2025. It combined three complementary models: Isolation Forest to detect broad outliers, Local Outlier Factor to identify local anomalies, and Exponentially Weighted Moving Average to track gradual shifts over time. These models were applied to the usual SCADA signals: differential pressure (ΔP), static pressure (P), and temperature (T). A single Composite Anomaly Index (CAI) was created to deliver a clear, easy-to-read score for the system's overall health. Of 288 data points, 14 anomalies (just under 5%) were flagged in which the CAI crossed the 0.5 threshold. The anomalies were not scattered randomly but appeared in clear temporal clusters. The strongest peaks occurred between late February and early March, then again from April into May, with CAI values reaching around 0.70–0.75. A milder set was observed in June–July, while August and September remained rock-steady with CAI scores mostly below 0.20. The sharp contrast between those calm periods and the intense spikes showed that the method effectively separated normal operation from real trouble. The integrated framework blended the three models to compensate for their individual weaknesses, capturing both sudden jumps and slow-creeping degradation patterns. Every flagged event aligned with physically expected field mechanisms, such as rising ΔP from fouling or steady baseline drifts from sensor wear. The framework runs entirely on existing SCADA data, with no need for labelled examples, making it cheap and straightforward to roll out at other metering stations.