DATA-DRIVEN PREDICTIVE MAINTENANCE IN GAS PLANTS: FAILURE PREDICTION WITH SUPPORT VECTOR MACHINES

Abstract

This study optimized data-driven predictive maintenance in gas plant using machine learning based Support Vector Machines (SVM) in predicting failures. The application software was developed to improve the operational efficiency and reliability of turbo-compressors in gas injection plants. The Gas injection plant produced below maximum capacity due to failure problems of the Turbo-compressors, these affected the targeted oil production negatively. The unavailability and unreliable gas plant led to revenue losses. Machine learning techniques using Support Vector Machines (SVM), was employed to develop the failure predictive application software. According to the findings, the Efficient Linear SVM model detected failures with a 99.5% true positive rate and classified non-failure events with a 99.9% classification precision. Although it showed a 0.5% false negative rate, the Boosted Trees model obtained a 99.5% true positive rate (TPR) for failure detection, underscoring the need for additional optimization and integration with ensemble approaches to reduce operational risks. Additionally, the SVM model demonstrated a minimum false negative occurrence and 99.9% classification precision for non-failure events. The outcomes of this study yielded a highly effective, computationally efficient machine learning-based application software capable of reliably predicting turbo-compressor failures. The study concluded that the developed application software is a powerful tool for predicting failures in gas injection plants, supporting decision-making processes, and enhancing operational safety. Recommendations for future works included refining existing models, exploring additional feature engineering techniques, and evaluating the robustness of the models under varying operational conditions