Classification and feature Regression for Multi-Phase Flow Regimes

Abstract

Detection and classification of flow-regimes are needed for improvements in the oil and gas sector, as well as in nuclear power plants. This thesis work presents a novel combination of convolutional neural networks applied to acoustic emissions from pipes containing multi-phase flow. A novel dataset is constructed from experimental data, and automatically labeled with video analysis. The proposed model classifies four distinct classes as well as performing a re- gression on the velocity and length of slugs appearing in the pipe. Both 2D and 1D transformations are tested in combination with different scaling methods. In addition the effectiveness of using a singular microphone is tested. The highest classification accuracy obtained on previously unseen data was 98.5%. The highest R2 score for slug velocity regression was 0.787 and the highest R2 score for slug length regression was 0.428. i