Vibration-Based Fault Detection in Centrifugal Pump with Machine Learning and Pre-Processing Filter

Lim, Kai Liang (2025) Vibration-Based Fault Detection in Centrifugal Pump with Machine Learning and Pre-Processing Filter. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text LIM KAI LIANG_FULL TEXT.pdf Restricted to Registered users only Download (5MB)

Abstract

This project investigates vibration-based fault detection in centrifugal pumps by assessing the impact of pre-processing filters on the accuracy of machine learning algorithms. It addresses the limitations of traditional methods such as physical monitoring, which frequently fail to detect early faults, resulting in high maintenance costs and downtime. Four filtering scenarios (Gaussian Median, Median Filter Only, Gaussian Filter Only, and No Filter) are used to pre-process vibration signals that were collected under normal, looseness, and leakage situations in order to improve signal quality by lowering noise. An integrated Random Forest-LightGBM model is used to analyze these signals and categorize defects, analyzing the impact of pre-processing on diagnostic performance. At standard speed, the Gaussian-Median Filter outperforms Median Filter Only Scenario (87.8 %), Gaussian Filter Only Scenario(71.2 %) and No Filter Scenario (56.8 %) in terms of accuracy, according to the results. Frequency-specific analysis at 21.58 Hz (74 %) and 40.65 Hz (81 %) reveals lower accuracy and looseness recall (0.50 - 0.68) because of signal and noise issues. Although frequency-specific restrictions indicate the necessity for adaptive strategies to maintain resilience across operating conditions, this integrated system provides a dependable predictive maintenance solution for industries like as water treatment and oil and gas.