Cavitation diagnosis for water distribution pumps: An early-stage approach combing vibration signal-based neural network with high-speed photography

Abstract

As an essential component of municipal water distribution systems, centrifugal water pumps are of great significance to achieve urban sustainability. Cavitation is a common phenomenon in water pumps that cause energy inefficiency and mechanical failures. To prevent cavitation damages, an early-stage cavitation diagnosis approach combing vibration signal-based neural network with high-speed photography was proposed. An adaptive neural network was developed using vibration measurement and cavitation states predefined using high-speed cavitation images collected at an in-house laboratory pump system with transparent casings. The correlation among synchronized cavitation images, vibration signals and pump performance was investigated. Our analysis shows that the head-drop detection method commonly used in the industry greatly underestimated the damage of cavitation with the fact that a 3% head drop corresponded to a cavitation intensity of 42.1%. Both the number of predefined cavitation states for training and the structure of neural networks greatly affected diagnosis accuracy and computing load. A two-stage ANN structure with eight cavitation states displayed the best performance with a much faster training speed compared with common shallow learning methods and consistent diagnosis accuracy of over 95% in real time. A water-energy-carbon nexus model was built to demonstrate provincial energy-saving potentials associated with cavitation prevention in China.