Machine Learning Based Flow Regime Identification using Ultrasonic Doppler Data and Feature Relevance Determination

Abstract

Machine learning approaches are now being used in the process industries to objectively classify two-phase flow regimes in pipes such as into bubbly, slugging, churning, or annular flow. However, they lack explainability, which makes the results difficult to be trusted by plant operators in the field. This paper presents a complete machine learning workflow for flow regime identification with improved explainability through the use of virtual flow regime maps and feature relevance determination. The workflow was applied on an S-shape riser fitted with a non-intrusive ultrasonic Doppler sensor. Statistical features were first extracted via discrete wavelet transform (DWT). An improved virtual flow regime map was then constructed by projecting the DWT features onto 2-D space using kernel principal components analysis (KPCA). Within KPCA, anisotropic kernels are proposed in this study (referred to as Aniso-KPCA) to reveal which input DWT features have the best flow-distinguishing power. Kernel parameters in Aniso-KPCA were searched by maximizing the goodness of 2-D clustering of data points using the genetic algorithm. Finally, the support vector machine (SVM) emerged as the recommended classifier of flow regimes on the basis of repeatability and accuracy, as compared to other models such as neural networks, naive Bayes, k-nearest neighbors, and ensembled trees. Overall, the proposed workflow is a step towards improving the explainability of future machine learning based flow regime classifiers.