Monitoring the condition of monoblock centrifugal pumps using vibration analysis and WiSARD classifier

• Published in: Journal of the Brazilian Society of Mechanical Sciences and Engineering 2025-01, Vol.47 (1)
• Main Authors: Shah, Om Rajesh, Sridharan, Naveen Venkatesh, Ramasamy, Sakthivel Nanjagoundenpalayam, Vaithiyanathan, Sugumaran
• Format: Article
• Language: English
• Subjects: Accelerometers; Accuracy; Algorithms; Bleaches; Centrifugal pumps; Classification (centrifugal); Condition monitoring; Data acquisition; Decision trees; Drift och underhållsteknik; Engineering; Fault detection; Fault diagnosis; Impellers; Mechanical Engineering; Monoblock centrifugal pumps; Network reliability; Neural networks; Operation and Maintenance Engineering; Parameter identification; Piezoelectricity; Statistical features; Task complexity; Technical Paper; Vibration analysis; Vibration monitoring; Weightless neural networks (WNN); WiSARD
• ISSN: 1678-5878; 1806-3691; 1806-3691
• DOI: 10.1007/s40430-024-05340-9

Enhancing the operational efficiency of centrifugal pumps across diverse industries necessitates a meticulous approach to fault diagnosis. This study presents an innovative methodology for detecting faults in centrifugal pumps, utilizing the revolutionary Weightless Neural Network (WNN). The primary aim of this method is to enhance the accuracy and reliability of fault detection by encompassing various fault categories and employing a comprehensive set of features. The fault diagnosis framework focuses on six specific conditions within the centrifugal pump system: cavity, faulty bearing, faulty bearing and impeller, faulty impeller, faulty seal, and healthy conditions. The initial step involves capturing vibration signals from these conditions using a piezoelectric accelerometer coupled with a data acquisition system. Subsequently, these signals undergo thorough processing to extract crucial insights facilitated by descriptive statistical features. To boost diagnostic accuracy, the study employs the J48 decision tree algorithm to identify the most significant features. In this context, the weightless neural network takes the form of a WiSARD classifier, and its parameters, namely tic number, bleach confidence, map type, bleach flag, bleach step, and bit number, are systematically fine-tuned to identify the optimal configuration for the statistical feature set. After rigorous experimentation, the results demonstrate that optimal hyperparameter tuning achieves an impressive classification accuracy of 99.17%. This comprehensive analysis not only enhances understanding of potential operational pitfalls but also facilitates proactive maintenance, preventing catastrophic failures. The WiSARD classifier emerges as the cornerstone of this innovative methodology, highlighting its efficacy in handling intricate fault classification tasks. Its adaptability and learning capabilities in discerning complex fault patterns position it as a formidable tool for the precise diagnosis of faults in centrifugal pumps.