Convolutional Neural Networks and Feature Fusion for Flow Pattern Identification of the Subsea Jumper

Abstract

The gas–liquid two-phase flow patterns of subsea jumpers are identified in this work using a multi-sensor information fusion technique, simultaneously collecting vibration signals and electrical capacitance tomography of stratified flow, slug flow, annular flow, and bubbly flow. The samples are then processed to obtain the data set. Additionally, the samples are trained and learned using the convolutional neural network (CNN) and feature fusion model, which are built based on experimental data. Finally, the four kinds of flow pattern samples are identified. The overall identification accuracy of the model is 95.3% for four patterns of gas–liquid two-phase flow in the jumper. Through the research of flow profile identification, the disadvantages of single sensor testing angle and incomplete information are dramatically improved, which has a great significance on the subsea jumper’s operation safety.