Prediction of cavitation using machine learning techniques on centrifugal pump

Abstract When the local static pressure drops below the fluid’s vapour pressure at that particular temperature, a dynamic process known as cavitation takes place in the liquid. Due to cavitation, the performance (total head and flow rate) of the pump will deteriorate, and hence the vibration and noise will develop. Both the performance and cavitation investigations are carried out for a low-specific-speed pump as per the testing standards. The results obtained during cavitation studies are extracted to train the machine learning algorithms. The flow rate, speed, torque, suction head, and delivery head are considered as input variables for machine learning, with total head, Net Positive Suction Head (NPSH), efficiency, and noise magnitude being output variables. The current studies employ four machine learning techniques: random forest, support vector machines, extreme gradient boosting, and linear regression. In the analysis, notable performance was observed for the random forest and extreme gradient boosting algorithms among others consistently demonstrated superior predictive capabilities for the output parameters of the pumps. For linear regression, random forest, and extreme gradient boosting, the coefficient correlation values are around one for all output parameters except for noise. Similar observation is depicted for coefficient of determination and normalized root mean squared error. These ensure the level of noise during cavitation is very dynamic. These findings will enable the pump user to incorporate AI techniques for the diagnosis of centrifugal pump operation with reference to the nominal flow rate (best efficiency flow). Through this technology, preventive and predictive maintenance of pumps can be adapted with less downtime in process industries, power plants, chemical industries, oil and gas refineries, and so on.