Artificial Intelligence-Based Diagnosis of Hydraulic Centrifugal Pumps Using Vibration Sensor: Feature Extraction and Failure Classification

Hydraulic pumps are deployed in a wide range of critical applications where reliability is a must. Condition monitoring of the pumps can reveal early faults and/or operational misuses to avoid expensive and dangerous breakdowns and to improve availability and reliability. Vibration-based monitoring techniques are well-known in this context. Methods based on Support Vector Machine (SVM), Artificial Neural Networks (ANNs), and K-Nearest Neighbors (KNN) are particularly well-established, and promising results have been reported using these techniques. In this study, SVM, ANN, and KNN are applied to detect and classify 6 different anomalies for a hydraulic centrifugal pump case study. The performances are evaluated using fast Fourier transform values of the vibration signals collected from a real-life setup. Different settings and methods for data analysis and feature extraction such as principal component analysis and spectral analysis are considered to achieve suitable configurations for the classifiers. The analysis takes into account vibration sensors placed on the housings of a pump and an electrical motor to examine the impact of sensor location on the performance of classifiers. Among the three methods, the ANN method with a 59-100-7 structure trained using pump vibration data achieves the best performance offering >99% and >98% of accuracy in the detection and classification of the anomalies, respectively.