Vibration-based multiphase-flow pattern classification via machine learning techniques

Abstract

Multiphase flow-induced vibration (FIV) is a complex problem since liquid and gas phases can present different spatial distributions, denoted as flow patterns, with distinct structural loading characteristics. This work investigates a non-intrusive measurement approach based on strain sensors and accelerometers that directly measure structural FIV. Nevertheless, a connection between the acquired data and the flow pattern cannot be directly derived since it depends on the structural dynamics. Therefore, a closed theoretical framework for such a fluid–structure interaction problem remains an open question. This work uses machine learning-based classifiers to classify two-phase flow patterns based on typical frequency domain features extracted from the data. This method has been validated against a measurement campaign that included 120 different flow conditions clustered into three flow patterns: Bubbles, Churn and Intermittent. The analyses revealed that these flow pattern classes are non-linearly separable. Moreover, regarding the bubbles and intermittent flow patterns, the results include a list of the most appropriate classifiers, the minimum required instrumentation for classification, and a physical interpretation of the suitability of the selected sensors and classifiers. However, due to the limited number of churn flow pattern data, the results regarding this pattern might indicate insights instead of conclusive results.