Development of a deep rectifier neural network for fluid volume fraction prediction in multiphase flows by gamma-ray densitometry

Abstract

This paper presents a novel methodology for volume fraction predictions in multiphase flow meters in offshore petroleum industries using gamma-ray densitometry. The algorithm that interprets spectra recorded at detectors is based on the new architectures of deep neural networks (deep learning). In this study, an 8-layer deep rectifier neural network (DRNN) has been used, running on a GPU-based parallel framework in order to increase the measuring precision of volumetric percentage of gas-oil-water system. The adaptive learning rate optimization algorithm (Adam) was used for DRNN training. All spectra recorded at detectors were normalized and used as the inputs of the DRNN. The detection system uses two NaI(Tl) detectors and appropriate narrow beam geometry, comprised of a (59.54 and 662 keV) dual-energy gamma-ray source and a dual-modality in order measure transmitted and scattered beams. The theoretical models consider a static oil-water-gas annular flow regime that were developed using MCNP6 code, which was used to provide training, validation and test data for the DRNN. The response-function of a real NaI(Tl) detector is also considered on the mathematical model. The results showed the root mean square error below than 0.8 for the fluids investigated in this study.