Blockage Diagnosis in Centrifugal Pump Using Stacking Ensemble Model

Blockages in the centrifugal pump affect the pump’s performance. It reduces the pump’s flow rate and increases noise, vibration, and overheating if left unnoticed for a prolonged time. Therefore, detecting these blockages in the pump is of utmost importance. This study focuses on detecting blockages in suction, discharge and simultaneous blockages in both lines at three severity levels using discharge pressure signals only. Initially, 12 statistical features extracted from the acquired pressure signal are normalised using the Z-score normalised method and their importance is prioritized using the minimum redundancy maximum relevance algorithm. A sensitivity analysis is carried out to select impactful features using a grid search cross validation method on Support Vector Machine which achieved peak accuracy when trained on the top eight features prioritized based on minimum redundancy maximum relevance algorithm. These eight features are then used to train the base learner of the stacking ensemble model namely, SVM, Random Forest, and XGBoost with optimized parameters using the GS algorithm. The Logistic regression serves as the meta-learner that integrates with the base learner by training on their average probabilities. The results show that the stacked ensemble model outperforms individual base learner classifiers with an accuracy of 93.63%. Thus, the stacked ensemble model proves its effectiveness in diagnosing blockages with enhanced reliability and precision.