Vibrothermal signal-based fault detection in 3D-printed polymer pump impellers

Abstract

Given the advancements in rapid prototyping and the increasing adoption of polymer materials in the fluid pump industry, effective fault diagnosis of centrifugal pump (CP) impellers is crucial. This study presents a novel hybrid Convolutional Neural Network with a Random Forest (CNN-RF) framework for fault detection in a 3D-printed PA6 polymer impeller, addressing the unique vibration characteristics of polymers compared to metal impellers. This study extends prior research by incorporating a comparative analysis of multinomial logistic regression, random forest, decision tree, support vector machine (SVM), and gradient boosting machine (GBM), demonstrating their strengths and trade-offs in predictive maintenance applications. An experimental setup employing vibration and temperature sensors served to gather data under three impeller conditions: healthy, one missing blade, and impeller with deformed disk. Advanced signal processing techniques facilitated feature extraction, and precision, recall, F1 index, area under the ROC curve, mean absolute error (MAE), and mean squared error (MSE) assessed model performance. Furthermore, stratified K-fold cross-validation, calibration curves, confusion matrices, and PCA served to ensure model robustness. Results demonstrate that Hybrid CNN-RF achieves the highest accuracy (99.88 %), outperforming conventional classifiers, while Random Forest and GBM also show strong predictive performance. The study further highlights the correlation between impeller faults and motor temperature variations, reinforcing the importance of thermal monitoring in fault diagnosis. This research advances predictive maintenance strategies by introducing a scalable and highly reliable machine learning-driven framework, bridging a gap in the diagnostics of polymer-based pump components.