Prediction of flow blockages and impending cavitation in centrifugal pumps using Support Vector Machine (SVM) algorithms based on vibration measurements

• Published in: Measurement : journal of the International Measurement Confederation 2018-12, Vol.130, p.44-56
• Main Authors: Panda, Asish Kumar, Rapur, Janani Shruti, Tiwari, Rajiv
• Format: Article
• Language: English
• Subjects: Accelerometers; Algorithms; Axial flow pumps; Binary classification; Casing; Cavitation; Centrifugal pump; Centrifugal pumps; Centrifuges; Computer simulation; Condition monitoring; Experimentation; Fault diagnosis; Feature extraction; Flow blockage; Inlet pipes; Machine learning; Multiclass classification; Predictions; Support Vector Machine (SVM); Support vector machines; Training; Vibration; Vibration monitoring
• ISSN: 0263-2241; 1873-412X
• DOI: 10.1016/j.measurement.2018.07.092

[Display omitted] •Fault diagnosis of centrifugal pumps using vibration signals.•Blockage severity and cavitation inception detection.•Support vector machine for multi fault classification.•99.4% accuracy in cavitation detection. The present work concentrates on the vibration based condition monitoring and fault diagnosis of centrifugal pumps. Two types of interrelated faults, i.e. flow blockages in the inlet pipe and impending bubble formation in the pump are considered. For the fault diagnosis and classification in the pump at varied speeds, a machine learning algorithm called, the Support Vector Machine (SVM) is utilized. Centrifugal pump is mounted on the Machine Fault Simulator (MFS™) set-up for the purpose of experimentation. Two tri-axial accelerometers, one on the pump casing and another on the bearing housing, are used to extract the vibration signals. Vibration signatures are taken at different flow blockages (0%, 16.7%, 33.3%, 50% and 66.6% of blockage) and at the start of bubble formation (inception of cavitation). Several statistical features are extracted from time domain vibration signal and fed to the SVM algorithm for training and testing. Standard deviation alone proves to be better than any other feature in this domain, and for this application. SVM parameters, including γ and C are optimally chosen. The ratio of training and testing data is also optimized. Binary fault classification offered better prediction accuracy for all blockage conditions over the multi-class fault classification. Moderately higher prediction accuracy in the multi-class fault classification (different level of blockages) has been found, when the training and the testing is done at higher rotational speeds. It has also been observed that the impending bubble formation could be very accurately predicted at higher rotational speeds.