Deep Learning Reconstruction Method For Electrostatic Tomography

Electrostatic tomography(EST)images by sensing the charge carried by solid particles at the electrodes.The number and intensity of the detected signals are much smaller than other actively excited electrical tomography techniques(e.g.,capacitance tomography,electromagnetic tomography,etc.).Therefore,it is more challenging to reconstruct high-quality images.The deep learning method can avoid the linear approximation process of most traditional imaging algorithms,so it can obtain higher image quality than traditional algorithms under the condition of less monitoring information.However,how to apply deep learning method to EST and optimize this method according to the characteristics of EST are still urgent problems to be solved.The main content of this thesis is as follows:1.The sensitivity-based electrostatic tomography algorithm is implemented,and the traditional Landweber algorithm is improved by setting a more reasonable initial iterative value to alleviate the existing local optimal problem,so as to improve the imaging quality of the central area of the pipeline.Simulation and experimental results show that the integrated reconstruction quality of the improved Landweber algorithm is better than that of the traditional Landweber algorithm,but the overall image quality of these sensitivitiy-based electrostatic tomography algorithms is still low.2.In order to solve the problem of lack of charge distribution and velocity distribution dataset for deep learning,charge distribution dataset is established through conventional static simulation,and dynamic velocity distribution dataset is established through a coupled simulation based on gas-solid flow field and static electric field.A total of 35,000 samples were collected in the charge distribution dataset,and a total of 5071 samples were collected in the velocity distribution dataset.3.In order to solve the problems that the deep learning-based charge distribution reconstruction method uses fewer network layers and the traditional loss function lacks prior information of EST,a prior knowledge based ResNet(PK-ResNet)image reconstruction method is proposed.The model is divided into three parts,that is,initial feature extraction module,residual feature extraction module and image reconstruction module.In the training process,the positive problem model of EST as the prior information of EST is encapsulated as a loss function,so as to improve the performance of the model.Simulation and experimental results verify the feasibility of the proposed model for charge distribution reconstruction,and the proposed loss function can improve the accuracy of reconstructed images.4.In order to apply deep learning method to reconstruct velocity distribution and the fusion of particle position information and velocity information,a Res+UNet image reconstruction method is proposed,which contains two ResNet sub-networks and one UNet sub-network.Two ResNet sub-networks are used to extract particle location information and velocity information respectively,and the UNet sub-network integrates the location information and velocity information to obtain velocity distribution images.Simulation and experimental results show that the Res+UNet model can reconstruct the state change of solid particles from deposition to suspension with the increase of apparent gas velocity,and the reconstructed velocity value increases with the increase of apparent gas velocity.