Image reconstruction algorithm for a lstm-based electrical capacitance tomography system

Abstract

The application relates to an image reconstruction method of an LSTM-based electrical capacitance tomography system, which comprises the following steps: (1) constructing an electrical capacitance tomography system model, obtaining a sensitivity field of the electrical capacitance tomography system through Comsol finite element simulation software, and obtaining electrical capacitance vectors under different positions and different shapes of an object field distribution; (2) obtaining a reconstructed image through a Landweber algorithm; (3) determining a ''time sequence'' sample according to ''rows'', and obtaining an image reconstruction result through an LSTM neural network model; (4) determining a ''time sequence'' sample according to ''columns'', and obtaining an image reconstruction result through an LSTM neural network model; and (5) fusing two image results to obtain a final image reconstruction result. Through the above method, the reconstructed image quality is improved, and the reconstructed medium distribution is closer to the real distribution.

Description

Technical Field

[0001] This invention relates to an image reconstruction method for an LSTM-based capacitance tomography system, belonging to the field of process tomography technology. Background Technology

[0002] Capacitive tomography (ECT) is a type of electrical tomography, a process imaging technique developed in the mid-to-late 1980s for detecting parameters in multiphase flow. It is an important process tomography technique based on medical CT technology and developed alongside computer and sensor technologies. The principle of ECT can be described as follows: if two substances with different dielectric constants are mixed in a pipe or container, changes in the components and their distribution will cause changes in the equivalent dielectric constant of the mixture, thus changing its measured capacitance value. Using a multi-electrode array capacitance sensor, the combination of electrodes can provide multiple capacitance measurements reflecting the dielectric constant distribution of the mixture. Using this as projection data and employing a specific image reconstruction algorithm, an image reflecting the medium distribution in a measured area of ​​the pipe or device can be reconstructed. ECT is an intelligent real-time process parameter detection system that non-invasively acquires two-dimensional and three-dimensional visualization information of the cross-section or spatial condition of pipes or containers in complex industrial processes. This provides an effective approach for detecting parameters with complex and variable characteristics that are difficult to detect using conventional methods. Currently, a great deal of work has been done on ECT technology both domestically and internationally, covering fields such as chemical engineering, petroleum, metallurgy, energy, power, light industry, and nuclear energy, and has yielded rich scientific research results in these fields.

[0003] Because ECT (Electronic Contrast Image Reconstruction) is still in its early stages and suffers from underdeterminism, ill-conditioned nature, nonlinearity, and "soft field" characteristics, current image reconstruction algorithms exhibit drawbacks such as low image accuracy and unsatisfactory edge reconstruction performance. Therefore, research on ECT image reconstruction algorithms is of great significance. Summary of the Invention

[0004] The technical problem this invention aims to solve is an image reconstruction method for capacitance tomography systems based on LSTM neural networks. This invention, while keeping the capacitance projection data unchanged, builds upon traditional image reconstruction methods by constructing a "time series" structure. It utilizes the long-term and short-term sequence memory capabilities of LSTM to complete the image reconstruction task of capacitance tomography, which to some extent overcomes the nonlinearity of capacitance tomography systems. This improves the quality of the reconstructed image, making the reconstructed medium distribution closer to the true distribution.

[0005] To solve the above problems, the specific technical solution of the present invention is as follows:

[0006] The image reconstruction method for an LSTM-based capacitance tomography system includes the following steps:

[0007] (1) Construct a capacitance tomography system model and obtain the sensitivity field of the capacitance tomography system using Comsol finite element simulation software; obtain the capacitance vector under different positions and shapes of object field distribution.

[0008] (2) Based on the sensitivity field and capacitance vector obtained in step (1), the Landweber algorithm is used to obtain the reconstructed image.

[0009] (3) Determine the "time series" samples by "row" and use the LSTM neural network model to obtain the image reconstruction results.

[0010] (3.1) Convert the reconstructed image obtained in step (2) into an input "time series" with time steps: connect the reconstructed image obtained in step (2) into a one-dimensional vector by connecting the beginning and end of the "rows"; and reorganize this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the input "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this input sequence.

[0011] (3.2) Convert the manifold image set by the Comsol finite element simulation software in step (1) into an output "time series" with time steps: Connect the manifold image set by the Comsol finite element simulation software in step (1) into a one-dimensional vector by connecting the beginning and end of the "rows"; rearrange this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the output "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this output sequence.

[0012] (3.3) Use the input sample set and output sample set obtained in steps (3.1) and (3.2) to train the LSTM neural network. After training, save the neural network model.

[0013] (3.4) During the inference phase, the neural network model is loaded, and the reconstructed image obtained by the Landweber algorithm is converted into a one-dimensional vector by concatenating the first and last rows. This one-dimensional vector is then rearranged into an N*M dimensional matrix, and these N rows of vectors are then fed into the LSTM neural network in sequence. Finally, the neural network outputs an N*M dimensional matrix. This N*M dimensional matrix is ​​then rearranged into a one-dimensional vector by concatenating the first and last rows. This vector can be rearranged into a two-dimensional square matrix n*n, which is the corresponding reconstructed image.

[0014] (4) Determine the "time series" samples by "column" and use the LSTM neural network model to obtain the image reconstruction results.

[0015] (4.1) Convert the reconstructed image obtained in step (2) into an input "time series" with time steps: connect the reconstructed image obtained in step (2) into a one-dimensional vector by connecting the first and last "columns"; and reorganize this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the input "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this input sequence.

[0016] (4.2) Convert the manifold image set by the Comsol finite element simulation software in step (1) into an output "time series" with time steps: Connect the manifold image set by the Comsol finite element simulation software in step (1) into a one-dimensional vector by connecting the first and last "columns"; reorganize this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the output "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this output sequence.

[0017] (4.3) Use the input sample set and output sample set obtained in steps (4.1) and (4.2) to train the LSTM neural network. After training, save the neural network model.

[0018] (4.4) During the inference phase, the neural network model is loaded, and the reconstructed image obtained by the Landweber algorithm is concatenated column-wise to form a one-dimensional vector. This one-dimensional vector is then rearranged into an N*M dimensional matrix, and these N rows of vectors are fed into the LSTM neural network in sequence. Finally, the neural network outputs an N*M dimensional matrix. This N*M dimensional matrix is ​​then concatenated column-wise to form a one-dimensional vector, which can be rearranged into a two-dimensional square matrix n*n, which is the corresponding reconstructed image.

[0019] (5) Fuse the two image results to obtain the final image reconstruction result: Add the corresponding pixels of the image reconstruction results obtained in steps (3) and (4), and take the average value as the pixel value of the fused image. This will give you the final image reconstruction result.

[0020] Beneficial effects of this invention:

[0021] This invention employs an LSTM neural network structure, which fully utilizes the nonlinear mapping characteristics of neural networks. It also leverages the sequential processing of natural language by LSTM neural networks, transforming the reconstructed image into a "time series" sentence with time steps. This effectively utilizes the correlation between image pixels to some extent, eliminating the need for a separate encoding library. Furthermore, this invention provides two different "time series" encoding schemes and fuses the reconstructed images obtained from the two schemes. This significantly improves the quality of the reconstructed image, making the reconstructed medium distribution closer to the true distribution. This invention also provides new approaches and methods for the research and application of capacitance tomography, possessing significant practical application value. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the working principle of an LSTM neural network.

[0023] Figure 2 This is a flowchart of the image reconstruction algorithm for an LSTM-based capacitance tomography system.

[0024] Figure 3 This is a schematic diagram illustrating the principle of image fusion based on LSTM neural network.

[0025] Figure 4 An example of a graphical display of an image after encoding. Detailed Implementation

[0026] like Figure 1 The diagram illustrates the working principle of an LSTM neural network. Figure 1 In the LSTM module, the cell state at the previous time step t-1 was C. t-1 The hidden state at the previous time t-1 was h. t-1 The input state at the current time t is X t , where h t It is also used as the output at time t. Here, X... t It is the current vector after image encoding, h t It is a current output vector of the LSTM neural network.

[0027] like Figure 2The diagram shows the image reconstruction algorithm flowchart for a capacitance tomography system based on an LSTM neural network. It mainly includes five stages: data preparation, LSTM neural network model training, data inference, image reconstruction, and image fusion. Data preparation primarily involves acquiring samples using Comsol, obtaining reconstructed image results using Landweber, and converting the reconstructed images obtained from Landweber and the Comsol manifold into "time series" samples according to row or column rules. LSTM neural network model training mainly includes sample partitioning, LSTM neural network structural parameters, neural network training, and model saving. Data inference mainly involves inferring from the input data using the above model to obtain the optimal result. Image reconstruction mainly uses the inference results to obtain the image reconstruction result required by the capacitance tomography system. Image fusion mainly fuses the image results encoded by "rows" and "columns" to obtain the final image reconstruction result of the capacitance tomography system.

[0028] like Figure 3 The diagram shown illustrates the image fusion principle based on LSTM neural networks. First, "time series" samples are determined by "row" and "column" respectively, and the corresponding image reconstruction results are obtained using the LSTM neural network model. Then, the two images are fused to obtain the final image reconstruction result of the capacitance tomography imaging system.

[0029] like Figure 4 The image shown is an example of the graphic display after image encoding. An original image (40*40) is divided into a one-dimensional vector, and then this one-dimensional vector is reconstructed into a two-dimensional matrix with 4 rows and 400 columns each. Here, to more intuitively display the encoding effect, each row is encoded into a two-dimensional image (20*20) for display.

[0030] The image reconstruction method of the capacitance tomography system based on LSTM neural network, taking the determination of "time series" samples by "row" as an example, includes the following steps:

[0031] (1) Data preparation stage

[0032] 1.1) Construct a capacitance tomography system model using Comsol finite element simulation software and obtain the sensitivity field of the capacitance tomography system; set up flow patterns such as single block, two blocks, three blocks, four blocks, five blocks, circulating flow, and laminar flow, and obtain capacitance vectors under different positions and shapes of the object field distribution under the above flow patterns as much as possible.

[0033] 1.2) Based on the sensitivity field and capacitance vector obtained in step 1.1), the Landweber algorithm is used to obtain reconstructed images with various manifold settings.

[0034] 1.3) Concatenate the reconstructed image obtained in step 1.2) into a one-dimensional vector by connecting the first and last rows. Rearrange this one-dimensional vector into an N*M dimensional matrix. These N rows can be considered as the input "time series" with time steps, and each row vector represents the "word embedding vector" of the words in this input sequence. In this way, the reconstructed image obtained in step 1.2) becomes an input "time series" with time steps. Correspondingly, perform the same processing on the manifold set in the Comsol finite element simulation software to obtain the output "time series" with time steps.

[0035] (2) Training phase of LSTM neural network model

[0036] 2.1) Determine the structural parameters of the LSTM neural network. Here, the LSTM used is the nn.LSTM module function in the PyTorch framework. The number of words in each input and output "time series" is 4, the number of features in the input matrix is ​​400, the number of hidden layer nodes is 500, the number of hidden layers is 3, and the dimension of the output is 400.

[0037] 2.2) Determine the training samples. Take 90% of the samples in the sample library as training samples and use the remaining samples as test samples.

[0038] 2.3) Train the LSTM neural network and save the neural network model after training is complete.

[0039] (3) Data reasoning stage

[0040] The neural network model is loaded, and the reconstructed image obtained by the Landweber algorithm is concatenated row by row to form a one-dimensional vector. This one-dimensional vector is then rearranged into an N*M dimensional matrix. These N row vectors are then fed sequentially into an LSTM neural network, and the optimal result is selected. Finally, an N*M dimensional matrix is ​​obtained. Here, N=4 and M=400.

[0041] (4) Image reconstruction stage

[0042] The image reconstruction stage mainly utilizes the optimal inference result obtained in step (3) to integrate the data, thereby obtaining the image reconstruction result required by the capacitance tomography imaging system. That is, this N*M dimensional matrix is ​​reorganized into a one-dimensional vector by concatenating the first and last rows, and then this vector is reorganized into a two-dimensional square matrix n*n, which is the corresponding reconstructed image. Here, N=4, M=400, and n=40.

[0043] (5) Image reconstruction algorithm of LSTM neural network for determining "time series" samples by "column".

[0044] The data preparation stage for determining the "time series" samples by "column" involves concatenating the first and last "columns" of the encoding process in step 1.3) above, while keeping other processes unchanged, thus obtaining another encoding result. The neural network model training stage for determining the "time series" samples by "column" is the same as step (2) above. In the data inference stage, the reconstructed image obtained by the Landweber algorithm is converted into a one-dimensional vector by concatenating the first and last "columns"; this one-dimensional vector is then rearranged into an N*M dimensional matrix, and these N rows of vectors are sequentially fed into the LSTM neural network to select the optimal result. The image reconstruction stage is consistent with step (4) above.

[0045] (6) Image fusion stage

[0046] By training an LSTM neural network using two different encoding methods—rows and columns—two different image reconstruction results can be obtained. Then, the corresponding pixels of these two reconstruction results are added together, and the average value is taken as the pixel value of the fused image, thus obtaining the final image reconstruction result.

[0047] The above description is merely a preferred embodiment of the present invention, and the method of the present invention can also be applied to other tomographic imaging systems such as electromagnetic tomography and electrical impedance tomography. It should be noted that those skilled in the art can make various modifications and improvements without departing from the principles of the present invention, and these modifications and improvements should also be considered within the scope of protection of the present invention.

Claims

1. An image reconstruction method for an LSTM-based capacitance tomography system, characterized in that, Includes the following steps: (1) Construct a model of the capacitance tomography system and obtain the sensitivity field of the capacitance tomography system through Comsol finite element simulation software; obtain the capacitance vector under different positions and shapes of object field distribution; (2) Based on the sensitivity field and capacitance vector obtained in step (1), the Landweber algorithm is used to obtain the reconstructed image; (3) Determine the "time series" samples by "row" and use the LSTM neural network model to obtain the image reconstruction results; The specific method is as follows: (3.1) Convert the reconstructed image obtained in step (2) into an input "time series" with time steps: connect the reconstructed image obtained in step (2) into a one-dimensional vector by connecting the beginning and end of the "rows"; and reorganize this one-dimensional vector into an N*M dimensional matrix. These N rows are regarded as the input "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this input sequence. (3.2) Convert the manifold image set by the Comsol finite element simulation software in step (1) into an output "time series" with time steps: Connect the manifold image set by the Comsol finite element simulation software in step (1) into a one-dimensional vector by connecting the beginning and end of the "rows"; rearrange this one-dimensional vector into an N*M dimensional matrix, and these N rows are regarded as the output "time series" with time steps. Each row vector is the "word embedding vector" representing the words in this output sequence. (3.3) Use the input sample set and output sample set obtained in steps (3.1) and (3.2) to train the LSTM neural network. After training, save the neural network model. (3.4) During the inference stage, the neural network model is loaded, and the reconstructed image obtained by the Landweber algorithm is connected from the beginning and end of the row to form a one-dimensional vector; this one-dimensional vector is rearranged into an N*M dimensional matrix, and then these N row vectors are fed into the LSTM neural network in sequence. Finally, the neural network outputs an N*M dimensional matrix. Then, this N*M dimensional matrix is ​​reorganized into a one-dimensional vector by concatenating the first and last rows. This vector can be reorganized into a two-dimensional square matrix n*n, which is the corresponding reconstructed image. (4) Determine the "time series" samples by "column" and use the LSTM neural network model to obtain the image reconstruction results; The specific method is as follows: (4.1) Convert the reconstructed image obtained in step (2) into an input "time series" with time steps: connect the reconstructed image obtained in step (2) into a one-dimensional vector by connecting the first and last "columns"; and reorganize this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the input "time series" with time steps, and each row vector is the "word embedding vector" representing the words in this input sequence. (4.2) Convert the manifold image set by the Comsol finite element simulation software in step (1) into an output "time series" with time steps: Connect the manifold image set by the Comsol finite element simulation software in step (1) into a one-dimensional vector by connecting the beginning and end of the "column"; rearrange this one-dimensional vector into an N*M dimensional matrix. These N rows can be regarded as the output "time series" with time steps. Each row vector is the "word embedding vector" representing the words in this output sequence. (4.3) Use the input sample set and output sample set obtained in steps (4.1) and (4.2) to train the LSTM neural network. After training, save the neural network model. (4.4) During the inference stage, the neural network model is loaded, and the reconstructed image obtained by the Landweber algorithm is connected end to end according to the "column" to become a one-dimensional vector; this one-dimensional vector is rearranged into an N*M dimensional matrix, and then these N rows of vectors are fed into the LSTM neural network in sequence. Finally, the neural network outputs an N*M dimensional matrix. Then, this N*M dimensional matrix is ​​reorganized into a one-dimensional vector by connecting the first and last columns. This vector can be reorganized into a two-dimensional square matrix n*n, which is the corresponding reconstructed image. (5) Fuse the two image results to obtain the final image reconstruction result.

2. The image reconstruction method of the LSTM-based capacitance tomography system according to claim 1, characterized in that, In step (5), the specific method is as follows: add the corresponding pixels of the image reconstruction results obtained in steps (3) and (4), and take the average value as the pixel value of the fused image, so as to obtain the final image reconstruction result.

Images