Predictive maintenance has emerged as a promising approach for improving the reliability and safety of pipeline and well infrastructure through continuous monitoring and data-driven decision-making. However, the development and evaluation of predictive maintenance systems in this domain are hindered by (i) the scarcity of labeled failure data, (ii) the nonstationary nature of sensor measurements, and (iii) the gap between predictive model outputs and actionable maintenance decisions.
This study presents a simulation-driven predictive maintenance framework that integrates synthetic multivariate sensor time-series data, machine learning based anomaly detection and remaining useful life (RUL) estimation, and an autonomous maintenance agent for risk- aware alert prioritization. The framework employs a configurable synthetic data generator to emulate realistic sensor dynamics and annotated failure events, enabling controlled and reproducible experimentation. Predictive models are trained to identify abnormal behavior and estimate degradation progression, while the autonomous agent supervises model outputs, monitors drift, schedules retraining, and translates predictive signals into prioritized maintenance alerts based on severity, urgency, and cost-consequence weights.
Simulation-based evaluation indicates that combining anomaly detection with RUL estimation yields complementary predictive evidence, and that agent supervision reduces false alarms while improving the relevance and prioritization of maintenance alerts. Overall, the study argues that a carefully designed, simulation-driven framework augmented by autonomous decision logic can support effective and interpretable maintenance decision- making for pipeline and well infrastructure under data-scarce and nonstationary conditions