Robust Machine Learning Framework for Predictive Maintenance in Gas Treatment Plants under Outlier-Influenced Operational Conditions

Abstract

The increasing complexity of gas treatment plants (GTPs) and the rising global demand for natural gas underscore the need for efficient and reliable maintenance strategies. Traditional maintenance approaches—reactive and preventive—have shown limitations in predicting failures accurately, often resulting in costly downtime and inefficient resource utilization. This study aims to evaluate the performance of selected machine learning (ML) models—Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), and K-Nearest Neighbors (KNN)—in predicting maintenance requirements in GTPs under data conditions involving outliers and the application of robust scaling techniques. Operational metering data from January 2019 to June 2024 were obtained from Total Energies EP Nigeria Limited. After preprocessing, feature engineering and selection were performed to prepare the dataset for training the models in Python’s Jupyter Notebook environment. The models were assessed based on accuracy, precision, recall, and F1-score. Results revealed that RF and DT models achieved an ideal 100% accuracy even in the presence of outliers and after applying robust feature scaling, although this level of perfection is unlikely in real-world conditions. KNN showed moderate performance with an accuracy of 89%, while SVM performed poorly with an accuracy of 47%, indicating its high sensitivity to outliers despite scaling. The findings confirm that tree-based algorithms such as RF and DT are more resilient and suitable for predictive maintenance tasks in noisy industrial datasets. Therefore, the study concludes that integrating robust scaling with tree-based ML models provides a more dependable solution for predictive maintenance in GTPs. It is recommended that GTPs adopt RF or DT models alongside robust data preprocessing techniques to improve fault prediction, reduce unplanned downtimes, and enhance operational reliability.