A Study on Comparison of Classification Algorithms for Pump Failure Prediction

Abstract

The reliability of pumps can be compromised by faults, impacting their functionality. Detecting these faults is crucial, and many studies have utilized motor current signals for this purpose. However, as pumps are rotational equipped, vibrations also play a vital role in fault identification. Rising pump failures have led to increased maintenance costs and unavailability, emphasizing the need for cost-effective and dependable machinery operation. This study addresses the imperative challenge of defect classification through the lens of predictive modeling. With a problem statement centered on achieving accurate and efficient identification of defects, this study’s objective is to evaluate the performance of five distinct algorithms: Fine Decision Tree, Medium Decision Tree, Bagged Trees (Ensemble), RUS-Boosted Trees, and Boosted Trees. Leveraging a comprehensive dataset, the study meticulously trained and tested each model, analyzing training accuracy, test accuracy, and Area Under the Curve (AUC) metrics. The results showcase the supremacy of the Fine Decision Tree (91.2% training accuracy, 74% test accuracy, AUC 0.80), the robustness of the Ensemble approach (Bagged Trees with 94.9% training accuracy, 99.9% test accuracy, and AUC 1.00), and the competitiveness of Boosted Trees (89.4% training accuracy, 72.2% test accuracy, AUC 0.79) in defect classification. Notably, Support Vector Machines (SVM), Artificial Neural Networks (ANN), and k-Nearest Neighbors (KNN) exhibited comparatively lower performance. Our study contributes valuable insights into the efficacy of these algorithms, guiding practitioners toward optimal model selection for defect classification scenarios. This research lays a foundation for enhanced decision-making in quality control and predictive maintenance, fostering advancements in the realm of defect prediction and classification.