One-dimensional structure reparameterized convolutional neural network for two-phase image reconstruction based on ERT

Abstract

Electrical resistance tomography (ERT) can be applied to two-phase flow pattern identification which is a key research direction for improving the operational safety of different industrial equipment systems with complex flow fields. Aiming at the existing problem that the traditional algorithm for defining flow patterns cannot accurately establish the mapping relationship between the measured voltage from ERT system and the two-phase flow conductivity distribution, a novel one-dimensional structure reparameterized convolutional neural network (1D-SRPCNN) algorithm for two-phase flow pattern image reconstruction based on ERT is proposed. First, finite element method and deep learning software framework are used to build dataset and train the neural network model respectively. Second, a deep residual network (ResNet) is used as the main network structure in the algorithm, and the one-dimensional multiscale feature extraction block (1DMSFE-Block) is improved by structural reparameterization. Then, multiscale convolution is introduced to 1DMSFE-Block for extracting features of different receptive field sizes and performing linear fusion, and the predicted two-phase flow conductivity pixel vector is obtained by the feature map passing with three fully connected layers. The results show that 1D-SRPCNN has high reconstruction performance, the average relative image error is 5.15%, the average correlation coefficient is 97.2%, and it has high anti-noise performance and generalization performance. Different experimental data also show that 1D-SRPCNN has high image reconstruction accuracy and efficiency. The research will provide important theoretical support for accurately identifying two-phase flow patterns in different fields.