Research On ECT Algorithm Based On Sparse Regularization And Deep Learning

Electrical Capacitance Tomography（ECT）,as a process Tomography technology,has the characteristics of non-invasion,high safety factor,fast response,convenient installation and low cost,suitable for industrial process detection visualization field.In order to deeply analyze the capacitance tomography technology,alleviate the under-qualitative problem and ill-conditioned phenomena in the process of capacitance tomography,this paper has launched several aspects of work:1.The sparse regularization algorithm is introduced to solve the inverse problem of ECT,and the basic concept and characteristics of sparse regularization are introduced.After comparing the applicability of the correlation sparse algorithm,the L_1/2regularization model was selected to replace the L₀regularization model.While avoiding solving the NP-hard problem,a more sparse solution is obtained,which improves the imaging quality and imaging speed.2.On the basis of the original semi-threshold iterative algorithm based on L_1/2regularization,L₂regularization term was added to deal with the phenomenon of non-smoothness at the origin.On this basis,the optimization item of the solution vector is added to alleviate the excessively strict calculation of the original semi-threshold iterative algorithm,which causes serious"cutting"marks in the flow pattern image.Finally,the imaging error of improved algorithm is reduced to 0.24,Correlation coefficient increased to0.92,and the imaging speed remains at 0.06s.3.In order to avoid ill-conditioned problems in the process of solving the inverse problem caused by the soft field effect,introduce the sequence prediction model 1D-CNN network in the field of deep learning,and use it in the process of ECT flow pattern identification.Introduced the basic progress of the combination of deep learning and ECT,and the basic concepts of the 1D-CNN network.By establishing a suitable 1D-CNN network to predict ECT image data,better imaging results than traditional algorithms are achieved.4.In order to solve the problem that a single neural network is difficult to capture complex and deep data features,the multi-head self-attention mechanism and the 1D-CNN network are combined to form a two-channel ECT image data prediction model.At the same time,for the purpose of improving the output weight on the convolution channel,the Squeez-and-Excitation network is added on the basis of 1D-CNN to form the MA-SE-CNN network structure.This method can simultaneously capture deep and representative feature information in tensor space and channel space.Compared with the traditional algorithm and the basic 1D-CNN,it more effectively solves the problem of the disappearance of important features when the ECT reconstruction image algorithm deals with the inversion of complex flow pattern.5.The imaging effect and data index of the traditional algorithm,the improved semi-threshold algorithm based on sparse regularization and the MA-SE-CNN neural network algorithm based on deep learning were compared,and analyze the effectiveness of each algorithm and the applicable scenarios of the algorithm.Experiments show that the two reconstruction algorithms proposed in this paper have significant improvements in imaging quality and imaging time compared with traditional algorithms.