ERT Flow Pattern Recognition And Image Reconstruction Based On Deep Learning

Electrical resistance tomography(ERT)is a cutting-edge visual detection technology,which is extensive used in petroleum,chemical and other industrial fields.Its advantages are low cost,no radioactivity and non-invasive.ERT technology first measures the field boundary voltage,and then reconstructs the medium conductivity distribution in the measurement area through a specific algorithm,so as to obtain the medium distribution in the field.In application,flow pattern recognition and image reconstruction algorithms play a vital role in obtaining the real situation inside the pipeline.However,ERT technology is affected by many factors,such as the "soft field" characteristics of the system and the ill condition of solving the inverse problem,which makes it difficult for the existing algorithms to obtain satisfactory recognition accuracy and reconstruction accuracy.To solve these problems,the methods of ERT flow pattern recognition and image reconstruction based on deep learning are studied and proposed in this paper.The main work is as follows:(1)In this paper,a multi classification ERT flow pattern recognition method based on pseudo image coding is proposed.Aiming at the problem of insufficient feature extraction of voltage data in direct flow pattern recognition method,a pseudo image coding method of ERT voltage data is designed.The method of transforming one-dimensional voltage signal into twodimensional pseudo image information as input samples not only ensures the integrity of the original data,but also enhances the data characteristics.After establishing the ERT voltage image database,the ERT flow pattern recognition network is constructed through convolution neural network to complete the multi classification recognition of ERT flow pattern.Experiments show that among the 27 types of flow pattern recognition,the average accuracy of ERT flow pattern recognition algorithm proposed in this paper is 98.74%,which is 10.64%higher than that of SVM,LSTM direct recognition method and CNN visual recognition method based on Newton-Raphson reconstruction.(2)An image reconstruction method of electrical resistance tomography based on adaptive neural module network is proposed in this paper.The database of gas-liquid two-phase distribution of various flow patterns is constructed,including boundary voltage,conductivity distribution and target image.The ERT flow pattern classifier with convolutional neural network structure is used to divide the flow patterns finely,so the VAE-GAN network corresponding to the flow pattern is selected for reconstruction.The variational auto-encoder structure is added to the generator to improve the generalization ability of the network to data features.In this reconstruction process,the nonlinear mapping relation can be obtained without calculating the sensitivity matrix.This method of adaptive selection and reconstruction network after classification can concentrate the phase transition region,avoid imaging blindness,and pay more attention to the details of flow pattern distribution.Through the experimental comparison of Tikhonov algorithm,Newton-Raphson and Landweber algorithms,as well as depth residual neural network and convolution neural network,the results show that the ERT image reconstruction algorithm proposed in this paper has high reconstruction accuracy,fast real-time performance and strong anti-noise ability.Compared with the two depth learning algorithms,the average relative error is reduced by 49.73%and the average correlation coefficient is increased by 28.87%.In this paper,based on ERT system,ERT flow pattern recognition and image reconstruction algorithms are designed through three deep learning models:convolution neural network,generative adversarial networks and variational auto-encoder.Experiments show that the proposed algorithm can improve the limitations and shortcomings of the existing algorithms effectively,and the recognition accuracy and reconstruction accuracy are improved significantly.This method is applied to ERT mine filling pipeline detection platform,which meets the needs of actual measurement and imaging,and has theoretical significance and practical application value.