Data-Driven Oscillation Root Cause Analysis in a Natural Gas Sweetening Unit: A Kalman Filter–Statistical Boundary Model Ensemble with a Three-Layer AI Agent Monitoring Platform

Abstract

Recurring pressure oscillations in amine-based natural gas sweetening units can be misinterpreted as solvent performance degradation, delaying mechanical investigation and prolonging unstable operation. This paper presents a practical, data-driven workflow for oscillation Root Cause Failure Analysis (RCFA) support in the amine regeneration section of a gas sweetening unit, using 5-minute DCS historian data from one of Saka Energi Indonesia's facilities. The workflow is organized into three layers. Layer 1 applies change-point detection and oscillation-band analysis to identify abnormal regimes in selected sensor tags for the regenerator, reflux drum, and flash drum. Layer 2 combines a Kalman Filter, used to detect abnormal controller- and valve-related behavior, with a Statistical Boundary Model that measures deviations from the normal operating envelope. The combined model achieved an Area Under the Curve (AUC) of 0.9032 under period-derived validation labels, outperforming the XGBoost baseline. An LSTM autoencoder was also used as an independent anomaly-checking layer to support the ensemble classification. Layer 3 applies Granger-based temporal precedence and lead-lag analysis, supported by an LLM-based RCFA agent, to convert model outputs into structured causal interpretations and corrective-action priorities. The analysis indicates that the instability is more consistent with control valve stiction than with a primary decline in amine performance. Abnormal valve-related behavior appeared approximately 90–100 minutes before the downstream machine learning model identified H₂S-related indication, showing that product-quality alarms act as lagging indicators rather than early warning signals. The observed pattern suggests persistent stick-slip valve movement, incomplete recovery after prior escalation, and propagation of upstream flow variation through thermal response into downstream regenerator differential-pressure oscillation.

The RSCFA output translated the analytics into three actionable engineering recommendations: inspect the suspected pressure-control valve, review upstream flow-control stability, and perform a short-term review of thermal coupling in the regeneration section. The proposed three-layer monitoring platform converts historian data into rule-based alerts, health indicators, and plain-language causal explanations to support maintenance and operations decisions.