Deep Learning Reconstruction Method For Resistance And Resistance/Ultrasonic Bimodal Fusion Tomography

As a rapid response visualization process detection technology,tomography has great application potential in multiphase flow process with multi-medium distribution.The existing tomographic image reconstruction methods are mainly based on approximate linear models,which cannot accurately describe the nonlinear relationship between the medium distribution and boundary measurement data,and it is difficult to essentially improve the quality of reconstructed image.In recent years,deep learning method,attracting widespread attention,has the characteristics of hierarchical representation of data,nonlinear modeling of different finite-dimensional spatial variables,automatic mining and feature extraction,which provides a new idea for solving the nonlinear relationship between medium distribution and boundary measurement data in tomographic image reconstruction of horizontal gas/water two-phase flow.Therefore,from the perspective of data-driven nonlinear modeling and model-driven feature learning,the quality of reconstructed images is improved by solving the nonlinear mapping relationship between boundary measurement and medium distribution,the selection of hyperparameters and prior information in the iterative image reconstruction method and the effective use of different modality information of resistance/ultrasonic fusion reconstruction.The main research work includes:（1）On the basis of introducing the research background and significance of tomography technology,the research status of image reconstruction methods of electrical and dual-modality tomography,the characteristics and existing problems of deep learning method in image reconstruction,the feasibility of using deep learning method to learn the nonlinear relationship between boundary measurement and medium distribution,and the fusion of deep learning and existing image reconstruction methods are analyzed.The research ideas of applying deep learning method to dual-modality fusion tomography are summarized,and the databases for training and testing of deep learning image reconstruction methods are created.（2）For the nonlinear mapping relationship between the boundary measurement voltage and the medium distribution in the image reconstruction of electrical resistance tomography,from the perspective of data-driven nonlinear modeling based on deep learning,a V-Net image reconstruction method is proposed.In the process of initial imaging,a five-layer fully connected network is used to convert the measurement information into pixel information in the image space.In the process of feature extraction,the self-learning method is used to mine effective features,and the under-determined image reconstruction problem is improved by increasing the feature dimensions.In the process of V-Net training,the loss function composed of the cross entropy output by the network,the cross entropy output by the initial imaging module and the L₂regularization term is used to constrain and supervise the learning of the reconstruction network.The simulation and experimental results verify the feasibility of the proposed method for image reconstruction.（3）For the problem of the sparse information flow and gradient flow in V-Net affecting the quality of reconstructed image,a densely connected VD-Net image reconstruction method is proposed.Four dense modules are used to promote the optimization of VD-Net by increasing the information flow and gradient flow of the image reconstruction network,and the image reconstruction model can better adapt to the nonlinear mapping relationship between boundary measurement data and medium distribution.The experimental results of discrete medium distribution show that the reconstruction quality of densely connected VD-Net is significantly improved compared with the V-Net,and the spatial resolution is better than existing image reconstruction algorithms.The dynamic experiment of horizontal gas/water stratified distribution verifies the practicability of VD-Net.（4）For the selection of hyperparameter and image prior information in iterative reconstruction method,from the perspective of model-driven feature learning,the Landweber iterative reconstruction network is proposed based on the Landweber iterative reconstruction method.The fully connected sub-network is used to learn the relaxation factor in the fidelity term for solving the problem of selecting the hyperparameters,and the convolution sub-network is used to learn the prior information of the image for solving the problem of extracting prior information.Then,the joint learning of hyperparameters and prior information can be realized in the process of network training.The experimental results of discrete medium distribution show that the Landweber iterative reconstruction network has higher image reconstruction accuracy than the existing image reconstruction methods.Compared with the VD-Net reconstruction network,it improves the reconstruction quality of complex distribution under noisy conditions.The reconstruction results of the flowing gas/water stratified distribution prove the practicability of the proposed Landweber iterative reconstruction network.（5）For the effective fusion of measurement information obtained by different sensitive principles in resistance/ultrasonic dual-modality image reconstruction,from the perspective of data-driven based deep learning fusion modeling,a dual-branch attention image reconstruction network is proposed.Two fully connected modules are used to transform the measurement information of different sensitive principles into image spatial information of the same dimension,which solves the heterogeneity of different modality information.In the process of feature extraction,the feature dimension is increased to alleviate the under-determined of image reconstruction,and the spatial pyramid pooling module is used to obtain multi-scale feature information.The attention mechanism judges the importance of different modal features to the reconstruction target according to the correlation between different modal feature information,and assigns different attention weights to the local features of different modalities to promote the effective fusion of dual-modality information and collaborative imaging.The experimental results show that the dual-branch attention image reconstruction network can improve the reconstruction accuracy and noise immunity compared with the dual-branch concatenation reconstruction network and the single-modality image reconstruction methods.