Electrical Resistance Tomography Reconstruction Algorithm Based On Convolution Neural Network And Transfer Learning

Electrical resistance tomography(ERT)has the advantages of no radiation,noninvasive and fast response.It has broad application prospects in the fields of industrial measurement and medical clinical monitoring.Image reconstruction is the key link of electrical resistance tomography system.The improvement of its imaging accuracy and real-time performance is of great significance to the production process monitoring.Improving the accuracy and speed of image reconstruction has always been a research hotspot in the field of ERT.In recent years,the depth learning method,which has attracted much attention,has the ability of self-learning and extracting features in different feature spaces.It is a new method to solve industrial process tomography.This subject takes the ERT image reconstruction problem as the research object,uses the convolutional neural network(CNN)to learn the nonlinear mapping relationship between the boundary measured voltage and the measured medium distribution,transforms the image reconstruction problem into a binary classification problem in the measured area,and improves the quality of image reconstruction.By adjusting the basic structure and training mode of the image reconstruction network,the generalization ability of the reconstruction network for the experimental data set is improved.The main research work includes:(1)Aiming at the establishment of ERT label database,COMSOL and MATLAB are used to establish the inclusion simulation database,including 40000 groups of simulation samples,and the system test experiment is designed to establish the measured database,including 3259 groups of experimental samples.The data set is divided into training set,verification set and test set by cross validation method.(2)Aiming at the nonlinear,ill posed and ill posed problems of inclusion image reconstruction,a 10 layer network structure including convolution layer,pooling layer and full connection layer is proposed,which is named Res2net4.The network uses the form of hierarchical residual to fuse the data feature information,and uses two full connection layers for image fitting.In the process of training the network,regularization parameters are added to prevent over fitting of the network,and the efficiency and robustness of network parameter updating are improved by setting exponential decay learning rate and moving average model.Finally,the experimental test set and simulation test set are used to verify the anti-noise and generalization ability of the algorithm.(3)Aiming at the limited generalization ability of image reconstruction network,CNN-SVM and CNN-Unet algorithms are proposed to compensate the impact of incomplete data sets on image reconstruction and improve the quality of image reconstruction by fine tuning the Res2net4 network structure.Aiming at the problem that ERT simulation data set and experimental data set have different data distribution,a neural network structure of CNN-DDC is proposed.By adding an MMD adaptation layer to the feature layer,the distance between simulation data domain and experimental data domain is calculated and added to the network loss function for training.Experimental results show that the reconstructed network has good generalization ability for experimental data sets.