Machine Learning-Driven Predictive Maintenance of Refinery Fired Heaters Using Industrial Process Data

Abstract

Machine learning (ML) provides a powerful approach for predictive maintenance of industrial equipment, enabling early detection of equipment degradation and supporting condition-based maintenance strategies. In Petroleum refineries, fired heaters are critical equipments where progressive fouling of heater coils requires periodic pigging to restore thermal efficiency and ensure safe operation. Conventional pigging schedules, based on operator experience or fixed interval maintenance policies, may lead to suboptimal outcomes, including premature maintenance interventions or unexpected performance deterioration. This study introduces a ML-based framework to forecast pigging requirements in refinery fired heaters for proactive maintenance planning. A representative heater was selected based on the operational severity and historical pigging frequency. Nine years of operating data were extracted from the refinery process historian using a custom-developed automated data acquisition interface. The collected data was pre-processed and transformed using domain-informed feature engineering workflow incorporating key process and combustion-related variables. Three supervised ML models—Logistic Regression, Random Forest and XGBoost were trained to predict pigging events upto 30 days in advance. Performance was assessed with standard classification metrics, supplemented by daily and event-level analyses to assess their practical utility for maintenance decision-making. The results demonstrate the viability of integrating long-term process data, engineering knowledge and ML techniques for early prediction of pigging requirements. Thus, offering a practical decision-support tool for refinery operators to improve maintenance planning, mitigate unplanned operational disruptions, enhance the reliability and availability of fired heater assets.