Predictive Maintenance for Pumps Based on Vibration Monitoring

Kam, Yik Hou (2025) Predictive Maintenance for Pumps Based on Vibration Monitoring. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text KAM YIK HOU_FULL TEXT.pdf Restricted to Registered users only Download (3MB)

Abstract

Industrial pumps play a critical role in various applications, particularly in sectors such as oil and gas, water management, and manufacturing, where pumps like Electric Submersible Pumps (ESPs) are essential for fluid transportation. Traditional preventive maintenance methods often lead to ine!ciencies and high costs due to their reliance on predetermined schedules or average failure rates, which can result in unnecessary maintenance or delayed fault detection. This study investigates the use of predictive maintenance as a more e!cient alternative by applying stacking ensemble model to analyze vibration data from publicity datasets A stacking ensemble model was developed and evaluated alongside other classifiers, with features derived from frequency-domain vibration signals. Initial results on the imbalanced dataset showed poor performance in minority fault classes such as misalignment and rubbing. To address this, SMOTE was applied, which significantly improved the model’s recall and F1-scores for these underrepresented classes. The final model obtained an overall test accuracy of 98.5% and AUC ROC score of 0.998 and demonstrated robust and consistent performance across all fault categories, as validated through classification metrics and simulated real-time testing on reserved data. The outcomes confirm that integrating vibration analysis with machine learning enhances fault detection capabilities, reduces unexpected downtime, and supports more e"ective, data-driven maintenance decisions for industrial pump systems.