Open Collections

A machine learning approach to classification of gas entrainment and impeller wear in centrifugal pumps

University of British Columbia

Centrifugal pumps are a fundamental part of fluid transport around the world. Consequently, they are also one of the world’s dominant energy consumers. The impacts of inefficient operation and undiagnosed wear are widely documented and can be disastrous environmentally, financially, and logistically. Though commercial tools and methods for monitoring pump performance are abundant, they are used infrequently in practice. This phenomenon derives from several factors, including monitoring systems’ poor scalability to function with large numbers of pumps, acquisition costs, the necessity for additional technical personnel, and stringent policies constraining process downtime. This thesis describes the development of an affordable, adaptable sensing method for classifying two conditions detrimental to centrifugal pump operation; gas entrainment and radial impeller wear. The method utilizes dynamic pressure measurements, collected at the pump discharge using a solitary, conventional pressure transducer. Decomposing these pressure fluctuations into a novel array of statistical features yields characteristic trends correlated to the target phenomena. These features are then used to train a series of machine learning algorithms, including multilayer perceptrons (MLP), support vector machines (SVM), and random forests, which are in turn used to characterize the target conditions using binary, multi-class, and regression methods. Dynamic pressure data for training and testing the classification algorithms is generated using simulated and experimental methods. The binary MLP model predicts gas entrainment exceeding a 2% void fraction of air with 90% accuracy, and radial wear exceeding 1.5% of the impeller diameter with 97% accuracy. The multi-class MLP classifies gas entrainment and radial impeller wear into severity classes spanning 1% increments with 62% and 82% success rates, respectively. The random forest regression model achieves a median prediction error of 0.44% for gas entrainment and 0.16% for impeller wear. The diagnostic system presented in this research is unique in that it is not conceived as a standalone tool for pump users, but rather a shared process to be trained and configured by the pump manufacturer, then implemented by the operators. In its envisioned application, the scope of the classified phenomena would be augmented by the manufacturer to capture a wide variety of pump performance characteristics.

Text

2021-02-05

Attribution-NonCommercial-NoDerivatives 4.0 International

http://hdl.handle.net/2429/77247

Mechanical Engineering

University of British Columbia

UBCV

http://creativecommons.org/licenses/by-nc-nd/4.0/