Image Reconstruction Based On Convolutional Neural Network For Electrical Resistance Tomography

Electrical Resistance Tomography(ERT)has the advantages of non-invasive,non-radiation and low cost,which has wide application potential in multiphase flow detection,biomedical imaging and other fields.Image reconstruction algorithm is a key of ERT system.The improvement of image accuracy and real-time performance is of great significance for monitoring and controlling production process,improving production stability and security.However,due to the image reconstruction process is a typical nonlinear,under-determined and ill-posed inverse problem.The existing image reconstruction algorithms are difficult to consider the accuracy and real-time at the same time.Therefore,it is necessary to explore a new theoretical framework for ERT image reconstruction to further improve the image quality.In this paper,the gas-water two-phase flow is taken as the measurement object.Convolutional Neural Network(CNN)algorithm is used to reconstruct the media distribution of a 2-D cross-section through end-to-end learning.The effective information in boundary measurements is extracted by the convolution algorithm.The extracted effective features are integrated and expressed by the fully connected neural networks.The image reconstruction problem is transformed into a binary classification problem of the imaging region to further improve the imaging accuracy.The specific research work is as follows:(1)In order to establish the database,COMSOL and MATLAB simulation is adopted to get the database of inclusion models and stratified models respectively.In order to conduct supervised learning effectively,each data includes 2-D cross-section media distribution and corresponding boundary voltage measurements.The media distribution is discretized and binary,and the corresponding boundary voltage measurements are normalized to eliminate system errors.10-fold cross validation is used to divide the database into training set and validation set.(2)In order to solve the non-linear and ill-conditioned problems in image reconstruction of inclusion models,a basic hierarchical framework with alternating convolution layers and pooling layers is proposed.To improve its performance,regularization,sliding average,learning rate attenuation and stochastic gradient descent modules are added.The algorithm has been applied to simulation data and experimental data respectively.Results show that the proposed algorithm has a good anti-noise and generalization ability.(3)To solve the problem of high under-determined and ill-posed of stratified models caused by electrode failure,a sparse batch normalized Convolutional Neural Network(SBN-CNN)algorithm is proposed,which alleviates the under-determined problem and gradient vanishing problem in training process by up-sampling,batch normalization and improving the direction of gradient updating.The anti-noise and generalization ability is further improved by data enhancement and the algorithm is validated in a real dynamic two-phase flow system.