An image reconstruction method for visual monitoring of multiphase flow

Abstract

An image reconstruction method for multiphase flow visualization monitoring. First step, ECT source image acquisition: Two-channel ECT source images are obtained respectively through the Tikhonov regularization and the Landweber iterative algorithm. Second step, wavelet image fusion: (1) Perform three-level wavelet decomposition on the two-channel Tikhonov regularization and Landweber iterative source images respectively; (2) Fuse the wavelet coefficients at each scale through Bayesian decision-making and maximum entropy thresholding; (3) Perform wavelet reconstruction on the fused coefficients through inverse wavelet transform. Third step, RCF image reconstruction: Input the wavelet-fused image into the RCF network for training and testing to obtain the ECT reconstructed image. The present invention has the advantages of reducing image artifacts, compensating for source image defects, and improving imaging quality, providing a new technical approach for the realization of multiphase flow visualization monitoring under complex flow patterns.

Description

Technical Field

[0001] The present invention relates to the technical field of multiphase flow measurement, and particularly to an image reconstruction method for multiphase flow visualization monitoring in electrical capacitance tomography (ECT). Background Art

[0002] Multiphase flow (or two-phase flow) widely exists in various petroleum industrial processes such as oil and gas exploration, petrochemical industry, oil and gas transportation, and environmental protection. The main existing forms of multiphase flow include gas-liquid, liquid-liquid, and gas-solid two-phase flows, etc. At present, the main multiphase flow measurement methods include orifice plate method or Venturi method, mass flow method, ray method, optical probe method, electrical method, tomography method, etc. Electrical capacitance tomography (ECT) has the advantages of non-radioactivity, non-contact, visualization, and low cost, and is considered to be one of the most attractive technologies in the field of multiphase flow measurement. However, due to the influence of factors such as multiphase flow pattern, non-linear field effect, and soft field characteristics, it is very difficult to accurately reconstruct ECT images. The classical ECT image reconstruction algorithms mainly include linear backprojection (LBP), Tikhonov regularization method, Landweber algorithm, conjugate gradient method, etc. These methods have good effects in solving the imaging problems of simple targets, while for complex flow patterns, their limited capabilities may lead to a serious decline in the quality of the reconstructed images.

[0003] Due to its advantages in image processing, speech recognition, signal processing, etc., the convolutional neural network (CNN) has become an effective method for ECT image reconstruction under complex manifolds. The patent application with the application number 202110331794.X discloses an ECT image reconstruction method based on a deep neural network, which calculates the mathematical model of ECT image reconstruction by using numerical analysis methods; classifies the data flow type through an improved AdaBoost ensemble algorithm; establishes a deep neural network model; trains the established deep neural network model; and finally uses the deep neural network to complete ECT image reconstruction. The patent application with the application number 202010109722.6 discloses an ECT image reconstruction method based on a deconvolution network, which uses the deconvolution network to extract the spatial features of the true distribution and can dynamically present the 2D image of the pipeline cross-section. However, the above patents all perform data training and image reconstruction for parts related to the neural network model structure such as feature extraction and network model training, without considering and dealing with the defects and deficiencies existing in the ECT source image itself. However, in the ECT image reconstruction of CNN, in addition to being related to the CNN model structure, the image reconstruction effect of CNN largely depends on the quality of the input source image. The methods disclosed in the existing patent applications directly input the ECT source image into the CNN network for image recognition and reconstruction. Due to the soft-field characteristics of ECT itself, the source images all have varying degrees of artifacts and phase interface deviations, which will lead to serious degradation of the quality of the reconstructed image and reduced adaptability of the method under complex manifolds. Summary of the Invention

[0004] Aiming at the problems existing in the visual monitoring of ECT multiphase flow, the purpose of the present invention is to provide an image reconstruction method for visual monitoring of multiphase flow, which optimizes the ECT source image through wavelet fusion and serves as the input of a more abundant convolutional feature (RCF) network, having the advantages of reducing image artifacts, compensating for source image defects, improving imaging quality, etc., and providing a new technical approach for the realization of visual monitoring of multiphase flow under complex manifolds.

[0005] To achieve the above purpose, the present invention adopts the following technical solutions:

[0006] An image reconstruction method for visual monitoring of multiphase flow, comprising the following steps:

[0007] The first step: ECT source image acquisition

[0008] Two channels of ECT source images are obtained respectively through the Tikhonov regularization and the Landweber iterative algorithm, as shown in formula (1) and formula (2):

[0009] g = (S T S + αI) -1 S Tλ (1)

[0010] where g is the grayscale value of the source image, S is the sensitivity matrix obtained through finite element simulation, S T is the transpose matrix of S, λ is the normalized measured capacitance value, I is the identity matrix, and α is the regularization parameter;

[0011] g k+1 = g k - α k S T (Sg k - λ) (2)

[0012] where g k+1 and g k are the (k + 1)-th and k-th iteration values of the grayscale of the source image, the number of iterations k = 100 - 500, S is the sensitivity matrix obtained through finite element simulation, λ is the normalized measured capacitance value, and the coefficient α k = 0.003 - 0.005;

[0013] Step 2: Wavelet image fusion

[0014] (1) Perform three-level wavelet decomposition on the two-channel Tikhonov-regularized and Landweber-iterated source images respectively;

[0015] (2) Fuse the wavelet coefficients at each scale through Bayesian decision-making and maximum entropy thresholding,

[0016] (3) Perform wavelet reconstruction on the fused coefficients through inverse wavelet transform;

[0017] The fusion in step (2) has the following specific rules:

[0018] Assume that the wavelet coefficients of the Tikhonov-regularized source image and the Landweber-iterated source image at the k-th scale are D k T and D k L , respectively. In addition, a composite coefficient of 1 / 2(D k T + D k L ) is constructed. Assume that the three events related to the above three coefficients are X T , X L and X T+L . Assume that X T , X L and X T+L satisfy the following formula:

[0019]

[0020]

[0021]

[0022] Then the wavelet coefficient D after the fusion of two channels k is as follows:

[0023]

[0024] where T opt K is the maximum entropy threshold of the wavelet at the k-th scale, obtained by the maximum entropy threshold theory, and θ 1 and θ 2 are correction coefficients;

[0025] Step 3: RCF image reconstruction

[0026] Input the image after wavelet fusion into the RCF network for training and testing to obtain the ECT reconstructed image.

[0027] Advantages of the present invention:

[0028] The image reconstruction method proposed by the present invention uses the Tikhonov regularized source image and the Landweber iterative source image to form a two-channel source image input, performs three-level wavelet decomposition on the source images of the two channels respectively, adopts a fusion rule combining Bayesian and maximum entropy to fuse the wavelet transform coefficients, and uses the RCF deep learning network for image reconstruction, which can achieve the purpose of reducing image artifacts, compensating for source image defects, and improving imaging quality. Compared with the ECT reconstruction method based on a simple deep learning network, the image quality is higher, and it has higher accuracy and stronger adaptability to multiphase flows under different flow patterns, and can better meet the needs of actual industrial production. Description of the drawings

[0029] Figure 1 is the structural diagram of the image reconstruction method of the present invention.

[0030] Figure 2 is the reconstructed image of the simulation data of the present invention under four simple flow patterns, where Figure 2 (a) is the core flow, Figure 2 (b) is the bubble flow, Figure 2 (c) is the stratified flow diagram, and 2(d) is the annular flow.

[0031] Figure 3 is the reconstructed image of the simulation data of the present invention under four relatively complex flow patterns, where Figure 3 (a) is the three-bubble model, Figure 3 (b) is the four-bubble model, Figure 3 (c) is the mixed model of stratified and bubble flows,Figure 3 (d) is a cross model.

[0032] Figure 4 This invention is about reconstructing images from experimental data under different flow patterns. Among them, Figure 4 (a) is downward eccentric flow, Figure 4 (b) is core flow, Figure 4 (c) is right eccentric flow, Figure 4 (d) is stratified flow with a gas holdup of 50%, Figure 4 (e) is stratified flow with a gas holdup of 81%, Figure 4 (f) is stratified flow with a gas holdup of 28%. Specific implementation

[0033] The following further elaborates on this invention in conjunction with the attached drawings.

[0034] Combined with Figure 1 , an image reconstruction method for multiphase flow visualization monitoring includes the following steps:

[0035] The first step: ECT source image acquisition

[0036] Respectively, through the Tikhonov regularization and Landweber iterative algorithms, ECT source images of two channels are obtained, as shown in Formula (1) and Formula (2):

[0037] g = (S T S + αI) -1 S T λ (1)

[0038] Among them, g is the gray value of the source image, S is the sensitivity matrix obtained through finite element simulation, S T is the transpose matrix of S, λ is the normalized measured capacitance value (obtained through the ECT system data acquisition unit), I is the identity matrix, and α is the regularization parameter, taking α = 0.2.

[0039] g k+1 = g k -α k S T (Sg k -λ) (2)

[0040] Among them, g k+1 and g k are the (k + 1)-th and k-th iteration values of the source image gray level, the number of iterations k = 100 - 500, S is the sensitivity matrix obtained through finite element simulation, S T is the transpose matrix of S, λ is the normalized measured capacitance value (obtained through the ECT system data acquisition unit), and the coefficient α k = 0.003 - 0.005.

[0041] Step 2: Wavelet image fusion

[0042] (1) Perform three-level wavelet decomposition on the Tikhonov regularization and Landweber iteration source images of the two channels respectively;

[0043] (2) Fuse the wavelet coefficients at each scale through Bayesian decision-making and maximum entropy thresholding;

[0044] (3) Perform wavelet reconstruction on the fused coefficients through inverse wavelet transform.

[0045] The fusion in step (2) has the following specific rules:

[0046] Assume that the wavelet coefficients of the Tikhonov regularization source image and the Landweber iteration source image at the k-th scale are D k T and D k L , respectively. In addition, a synthetic coefficient of 1 / 2(D k T +D k L ) is constructed. Assume that the three events related to the above three coefficients are X T , X L and X T+L . Assume that X T , X L and X T+L satisfy the following formula:

[0047]

[0048]

[0049]

[0050] Then the wavelet coefficient D k after fusion of the two channels is:

[0051]

[0052] where T opt K is the maximum entropy threshold of the wavelet at the k-th scale, obtained by the maximum entropy threshold theory, and θ 1 and θ 2 are correction coefficients. Take θ 1 = 0.8 and θ 2 = 0.1.

[0053] Step 3: RCF image reconstruction

[0054] The image after wavelet fusion is input into the RCF network for training and testing to obtain the ECT reconstructed image.

[0055] The method of the present invention is applied to the ECT system for visual measurement of multiphase flow. The ECT system consists of an ECT sensor, a data acquisition unit, and a PC. The reconstruction software formed by the method of the present invention is installed on the PC. During the measurement, the ECT sensor is connected to the data acquisition unit through a coaxial cable, and the measured capacitance data collected by the data acquisition unit is transmitted to the PC through a communication interface. Finally, the reconstruction software formed by the algorithm of the present invention and the acquired data are used for ECT image reconstruction to realize the visual measurement of multiphase flow. In the experiment, the multiphase flow under different flow patterns is reconstructed, and the media in the test pipeline are gas-liquid two-phase flow and gas-solid two-phase flow.

[0056] The following combines Figures 2 - 4 to illustrate the reconstruction image effect of the present invention under different flow patterns. The relative image error (IE), correlation coefficient (CC), and structural similarity index measure (SSIM) are used to quantitatively evaluate the quality of the reconstructed image.

[0057]

[0058]

[0059] Among them, g and are the gray level vectors of the real image and the reconstructed image respectively. and are the average values of the gray level vectors g and respectively. p is the number of pixels in the imaging area.

[0060]

[0061] Among them, μ g , σ g , and respectively refer to the local mean, standard deviation, and cross-covariance of the real image g and the reconstructed image . C 1 and C 2 are the regularization constants of image brightness and contrast.

[0062] To evaluate the feasibility of the algorithm of the present invention, a 12-electrode ECT model was established through finite element analysis (FEA), and image reconstructions of LBP, Tikhonov regularization, Landweber, CNN, and W-RCF algorithms were completed respectively. Among them, the pipeline was filled with oil-gas two-phase flow, and the relative dielectric constants of oil and gas were 3 and 1 respectively. The image reconstruction area was divided into a 64*64 grid. Case 1: Four flow patterns, namely core flow, bubble flow, stratified flow, and annular flow, were studied, and the reconstruction results are as Figure 2 shown. In addition, Tables 1, 2, and 3 give the quantification indexes IE, CC, and SSIM of the reconstruction image quality, and analyze the four flow patterns using LBP, Tikhonov regularization, Landweber, CNN, and W-RCF algorithms respectively. From Figure 2 it can be seen that under core flow, bubble flow, stratified flow, and annular flow, for the images reconstructed by CNN and W-RCF algorithms, the oil-gas boundaries are clear, without artifacts, and the image reconstruction quality is significantly better than that of LBP, Tikhonov regularization, and Landweber algorithms. As can be seen from Tables 1, 2, and 3, for the four flow patterns, the IE of CNN and W-RCF algorithms is smaller, the CC is larger, and the SSIM is also larger than that of other algorithms. Further considering the overall performance, for the four simple patterns, the reconstruction effect of the W-RCF algorithm is slightly higher than that of the CNN algorithm.

[0063] Table 1 IE performance indexes of different algorithms under simple flow patterns

[0064] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.325 0.213 0.266 0.211 0.082 b 0.309 0.236 0.194 0.109 0.112 c 0.202 0.206 0.428 0.156 0.073 d 0.393 0.254 0.176 0.102 0.095

[0065] Table 2 CC indexes of different algorithms under simple flow patterns

[0066] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.634 0.857 0.844 0.897 0.934 b 0.798 0.815 0.821 0.904 0.921 c 0.806 0.798 0.591 0.852 0.959 d 0.595 0.803 0.867 0.932 0.944

[0067] Table 3 SSIM indexes of different algorithms under simple flow patterns

[0068] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.256 0.621 0.497 0.699 0.732 b 0.352 0.624 0.625 0.757 0.751 c 0.316 0.338 0.235 0.468 0.695 d 0.358 0.459 0.597 0.633 0.642

[0069] To further verify the applicability of the algorithm to actual two-phase flow, simulation experiments were carried out under more complex flow patterns. Case 2: Four complex flow patterns, namely three-bubble model, four-bubble model, stratified and bubble flow mixed model, and cross model, were studied, and the reconstruction results of applying LBP, Tikhonov regularization, Landweber, CNN, and W-RCF algorithms are respectively as Figure 3 shown. In addition, Tables 4, 5, and 6 show the quantification indexes IE, CC, and SSIM of the reconstruction image quality of these algorithms for the four complex flow patterns. From Figure 3It can be seen that for the four more complex flow patterns, by using LBP, Tikhonov regularization, and the Landweber algorithm, the artifacts and noise in the reconstructed images increase significantly, and the edges of the two-phase flow become more blurred. For Figure 3 the flow patterns (a) and (c) in

[0070] Table 4 IE performance metrics of different algorithms under complex flow patterns

[0071] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.654 0.359 0.276 0.305 0.149 b 0.762 0.204 0.232 0.187 0.122 c 0.403 0.173 0.523 0.169 0.098 d 0.863 0.325 0.729 0.103 0.095

[0072] Table 5 CC performance metrics of different algorithms under complex flow patterns

[0073] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.506 0.717 0.803 0.794 0.884 b 0.452 0.795 0.824 0.847 0.891 c 0.793 0.828 0.581 0.846 0.929 d 0.343 0.813 0.417 0.903 0.914

[0074] Table 6 SSIM performance metrics of different algorithms under complex flow patterns

[0075] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.399 0.628 0.648 0.639 0.702 b 0.676 0.676 0.674 0.733 0.756 c 0.197 0.196 0.192 0.245 0.582 d 0.290 0.287 0.271 0.609 0.645

[0076] An indoor ECT two-phase flow static experiment was conducted to verify the effectiveness of the algorithm of the present invention in practical applications. To achieve ECT image reconstruction, the measured capacitance data under different flow patterns were collected by the data acquisition unit, and the sensitivity matrix was generated by the FEA software at the same time. In the experiment, eccentric flow, core flow, and stratified flow were studied, and the images reconstructed by LBP, Tikhonov regularization, Landweber, CNN, and W-RCF algorithms are as Figure 4 shown, and the IE, CC, and SSIM metrics of the reconstructed images are shown in Table 7, Table 8, and Table 9 respectively.

[0077] From Figure 4 it can be seen that there are obvious artifacts and unclear two-phase flow boundaries in the images reconstructed by using LBP, Tikhonov regularization, and the Landweber algorithm, while there are some artifacts in the CNN algorithm, such as Figure 4 (a) and (c) and Figure 4Edge missing in (d) and (e). Obviously, the image reconstructed by applying the W-RCF algorithm has no artifacts, the two-phase flow boundary is more obvious, and the image quality is higher than that of the LBP, Tikhonov regularization, Landweber, and CNN algorithms. It can be clearly seen from the data in Tables 7, 8, and 9 that the IE of the W-RCF algorithm is less than that of other algorithms. At the same time, the correlation between the image reconstruction result obtained by the W-RCF algorithm and the real image is greater than that of other algorithms. In addition, the structural similarity between the image reconstruction result obtained by the W-RCF algorithm and the real image is the largest, higher than that of other algorithms. Obviously, the image reconstruction quality using our algorithm is significantly better than that of the LBP, Tikhonov regularization, Landweber, and CNN algorithms. Therefore, from the overall performance, the W-RCF algorithm for ECT reconstruction has higher accuracy and stronger adaptability for multiphase flow than the LBP, Tikhonov regularization, Landweber, and CNN algorithms. From the simulation and experimental results, it provides an effective way for ECT image reconstruction.

[0078] Table 7 IE performance indicators of different algorithms under different flow patterns in experimental tests

[0079] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.887 0.212 0.563 0.199 0.117 b 0.705 0.187 0.182 ·0.094 0.088 c 0.716 0.232 0.406 0.199 0.134 d 0.445 0.341 0.304 0.224 0.121 e 0.523 0,327 0.391 0.286 0.148 f 0.421 0.300 0.286 0.258 0.114

[0080] Table 8 CC performance indicators of different algorithms under different flow patterns in experimental tests

[0081] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.303 0.805 0.604 0.864 0.892 b 0.488 0.838 0.821 0.906 0.939 c 0.394 0.810 0.591 0.834 0.876 d 0.745 0.799 0.712 0.859 0.946 e 0.712 0.789 0.801 0.810 0.877 f 0.791 0.834 0.831 0.868 0.901

[0082] Table 9 SSIM performance indicators of different algorithms under different flow patterns in experimental tests

[0083] Serial number LBP Landweber Tikhonov CNN W - RCF a 0.142 0.453 0.325 0.469 0.544 b 0.256 0.451 0.326 0.569 0.578 c 0.098 0.289 0.334 0.451 0.601 d 0.234 0.367 0.371 0.421 0.487 e 0.238 0.401 0.421 0.434 0.499 f 0.233 0.487 0.470 0.492 0.502 。

Claims

1. An image reconstruction method for multiphase flow visualization monitoring, characterized in that, it includes the following steps: The first step: ECT source image acquisition Two channels of ECT source images are obtained respectively through the Tikhonov regularization and the Landweber iteration algorithm; The second step: Wavelet image fusion (1) Perform three-level wavelet decomposition on the Tikhonov regularization and Landweber iteration source images of the two channels respectively; (2) Fuse the wavelet coefficients at each scale through Bayesian decision-making and maximum entropy threshold; (3) Perform wavelet reconstruction on the fused coefficients through inverse wavelet transform; The third step: RCF image reconstruction Input the wavelet-fused image into the RCF network for training and testing to obtain the ECT reconstructed image; The ECT source images of the two channels in the first step are shown in Formula (1) and Formula (2): g=(S Τ S + αI) -1 S Τ λ(1) where g is the gray value of the source image, S is the sensitivity matrix obtained through finite element simulation, S T is the transpose matrix of S, λ is the normalized measured capacitance value, I is the identity matrix, and α is the regularization parameter; g k+1 = g k - α k S Τ (Sg k - λ) (2) Among them, g k+1 and g k are the (k + 1)-th and k-th iteration values of the source image grayscale. The number of iterations k = 100 to 500. S is the sensitivity matrix obtained through finite element simulation. λ is the normalized measured capacitance value. The coefficient α k = 0.003 to 0.005; The fusion in step (2) of the second step is specifically as follows: Suppose the wavelet coefficients of the Tikhonov-regularized source image and the Landweber-iterative source image at the k-th scale are D k T and D k L , respectively. In addition, a synthetic coefficient of 1 / 2(D k T +D k L ) is constructed. Let the three events associated with the above three coefficients be X T , X L and X T+L . Assume that X T , X L and X T+L satisfy the following formula: Then the wavelet coefficient D after the fusion of the two channels k is as follows: Among them, T opt K is the maximum entropy threshold of the wavelet at the k-th level of scale, obtained by the maximum entropy threshold theory, θ 1 and θ 2 are correction coefficients.

Images