Objectives/Scope: A Multi-Phase Flow Meter (MPFM) performs Water Liquid Ratio (WLR) estimations using a dedicated sensor relying on one physical principle (e.g. electrical impedance). The accuracy of the WLR sensor might also be dependent on the flow properties. An approach based on Machine Learning techniques and multi-sensors data fusion has been implemented to enhance the reliability and accuracy of the WLR estimations in Multi Phase application using onboard sensor measurements of a MPFM. Methods, Procedures, Process: In order to improve the estimations of multi-phase applications, we exploit the availability of historical data collected with heterogeneous sensors; the underlying idea of the proposed approach is to exploit such data with Machine Learning supervised techniques to provide accurate measures. In this work we compare modern supervised learning approaches like Random Forest, Gradient Boosting techniques and Kernel methods. The proposed methods have a relatively simple form that can be deployed also in embedded applications. Results, Observations, Conclusions: In this work, we will show through extensive experiments that the proposed approach could improve the original estimations. The algorithms underlying the proposed approach have been trained using data collected at flow loops test facilities with different flow conditions. The best model has been chosen not only for its predictive performances, but also looking at the computational time needed to make a prediction and considering its robustness to outliers. As expected, depending on the dataset numerosity, the best performing model can change: we provide experimental results for various dataset sizes in order to help practitioners choose the best regression method depending on the available data numerosity. An additional considered aspect is the computational time of the various approaches, which may be a relevant characteristic to be evaluated before rolling out productive solutions. Novel/Additive Information: To increase the accuracy of MPFM, a sensor fusion technique that benefits from the many measurements collected by the MPFM, has been developed. Many different models have been compared on: prediction performances, confidence interval, robustness to outliers, execution time. The resulting model provides enhanced estimations equipped with confidence intervals that can be used for prediction quality assessment and associated risk management.
Sensor fusion and machine learning techniques to improve water cut measurements accuracy in multiphase application
Barbariol T.;Feltresi E.;Susto G. A.;Tescaro D.;Galvanin S.
2020
Abstract
Objectives/Scope: A Multi-Phase Flow Meter (MPFM) performs Water Liquid Ratio (WLR) estimations using a dedicated sensor relying on one physical principle (e.g. electrical impedance). The accuracy of the WLR sensor might also be dependent on the flow properties. An approach based on Machine Learning techniques and multi-sensors data fusion has been implemented to enhance the reliability and accuracy of the WLR estimations in Multi Phase application using onboard sensor measurements of a MPFM. Methods, Procedures, Process: In order to improve the estimations of multi-phase applications, we exploit the availability of historical data collected with heterogeneous sensors; the underlying idea of the proposed approach is to exploit such data with Machine Learning supervised techniques to provide accurate measures. In this work we compare modern supervised learning approaches like Random Forest, Gradient Boosting techniques and Kernel methods. The proposed methods have a relatively simple form that can be deployed also in embedded applications. Results, Observations, Conclusions: In this work, we will show through extensive experiments that the proposed approach could improve the original estimations. The algorithms underlying the proposed approach have been trained using data collected at flow loops test facilities with different flow conditions. The best model has been chosen not only for its predictive performances, but also looking at the computational time needed to make a prediction and considering its robustness to outliers. As expected, depending on the dataset numerosity, the best performing model can change: we provide experimental results for various dataset sizes in order to help practitioners choose the best regression method depending on the available data numerosity. An additional considered aspect is the computational time of the various approaches, which may be a relevant characteristic to be evaluated before rolling out productive solutions. Novel/Additive Information: To increase the accuracy of MPFM, a sensor fusion technique that benefits from the many measurements collected by the MPFM, has been developed. Many different models have been compared on: prediction performances, confidence interval, robustness to outliers, execution time. The resulting model provides enhanced estimations equipped with confidence intervals that can be used for prediction quality assessment and associated risk management.Pubblicazioni consigliate
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.