A 3D reconstruction method of combustion field with limited angles based on residual network and physical constraints

Abstract

The present invention discloses a method for three-dimensional reconstruction of a limited-angle combustion field using a residual network based on physical constraints, comprising the following steps: using one-dimensional projection data of a limited-angle combustion flame field projection map as input data; using the physical three-dimensional structural feature distribution of the combustion flame field as prior knowledge to constrain the flame field reconstruction result, and introducing it into the loss function as a regularization term; training the residual network based on the loss function; and establishing an optimized residual network model through training to achieve high-precision three-dimensional reconstruction of the limited-angle combustion flame field. The residual network model constructed by the present invention can achieve three-dimensional reconstruction of flame emission tomography of a flame combustion field under the condition of restricted projection angle. Compared with the reconstruction result of the traditional algebraic iteration algorithm (ART), the reconstruction result obtained by the residual network model based on physical constraints under the condition of restricted angle has higher reconstruction accuracy, and is superior in calculation time and has a faster reconstruction rate.

Description

Technical Field

[0001] The invention relates to the field of combustion flame three-dimensional imaging algorithms, and in particular to a residual network limited-angle combustion field three-dimensional reconstruction method based on physical constraints. Background Art

[0002] The three-dimensional measurement and real-time display of the combustion field play an important role in promoting combustion technology, such as advanced rocket technology and energy utilization technology. To achieve transient three-dimensional parameter reconstruction, it is necessary to first obtain multi-directional projections of the measured dynamic flame. In recent years, the flame chemiluminescence tomography (FCT) method has aroused widespread interest among researchers, and different forms of multi-directional projection acquisition systems have been established.

[0003] Theoretically, only when the number of collected projections satisfies the Nyquist sampling theorem can the ideal three-dimensional tomographic reconstruction of the test field be achieved. However, in actual industrial environments, including power generation boilers, gas turbines and aircraft engines, the combustion process is generally carried out in non-free space. Therefore, in view of the limited detection space provided, it is difficult to obtain uniformly distributed multi-directional projections within a range of 180° around the test field. Therefore, the present invention adopts an optical system based on a camera array arrangement, which can obtain high-density directional flame projections in a limited observation window as effective information of the test field for three-dimensional flame chemiluminescence tomography.

[0004] Since the obtained high-density directional projection is a projection under a limited observation window, the reconstruction result based on the traditional iterative reconstruction algorithm is still limited by insufficient projection angles, which will produce artifacts during the reconstruction process and affect the reconstruction quality.

[0005] In recent years, deep learning has flourished in many fields, bringing significant changes to CT (Computed Tomography) technology. The paper "A perspective on deep imaging" combines deep learning and CT technology to improve the quality of medical imaging. The paper "Low-dose CT via convolutional neural network." adopts a denoising method for low-dose CT images based on convolutional neural networks, which reduces artifacts in medical CT imaging and improves image quality. Compared with traditional iterative reconstruction algorithms, the convolutional neural network (CNN) method can remove reconstruction artifacts, protect image edges, and has great advantages in computing speed. In addition, the residual network (ResNet) has also been widely used because it can avoid the gradient vanishing and gradient explosion problems of deep networks. Therefore, based on the residual network, a higher accuracy training result can be obtained. The present invention provides a residual network limited angle combustion field three-dimensional reconstruction method based on physical constraints. The method introduces the physical characteristics of the field to be measured into the residual network training process. The optimized residual network model established has higher prediction accuracy and more reliable field distribution in the reconstruction of angle-restricted combustion fields. Summary of the invention

[0006] The purpose of the present invention is to provide a residual network finite angle combustion field three-dimensional reconstruction method based on physical constraints, so as to solve the problem of three-dimensional reconstruction of combustion flame emission tomography under finite angle incomplete data projection.

[0007] The technical solution to achieve the purpose of the present invention is: a three-dimensional reconstruction method of a residual network limited angle combustion field based on physical constraints, comprising the following steps:

[0008] Step 1: Capture N frames of continuous flame projection images through CCD cameras at M different positions of a flame emission tomography reconstruction device based on a camera array, where each frame is composed of M CCD cameras simultaneously capturing one flame projection image, N>500; extract the position containing the flame emission intensity projection in the M×N flame projection images, and use the one-dimensional projection data of the same row of projection images corresponding to the position as input data, where the pixel size of the input data is h×l, with a total of K images, K≥10000; divide the input data into a training set and a test set in a ratio of 8:2;

[0009] Step 2: Construct a three-dimensional flame loss function based on the three-dimensional physical structure distribution characteristics of the combustion flame field;

[0010] Step 3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the reconstructed three-dimensional field of flame tomography as label data into the residual network, and use the loss function to train the residual network to obtain an optimized residual network model;

[0011] Step 4: Input the test set into the residual network model, and then perform three-dimensional superposition reconstruction on the two-dimensional distribution predicted by the residual network model to obtain the flame tomography three-dimensional structure distribution.

[0012] Compared with the prior art, the present invention has the following significant advantages:

[0013] (1) The loss function based on the three-dimensional physical structure distribution of the flame proposed in the present invention can fully consider the three-dimensional structural characteristics of the flame combustion field, better measure the error between the predicted results and the actual results, and improve the robustness of the model;

[0014] (2) Compared with the traditional flame emission tomography algebraic iteration (ART) reconstruction algorithm, the algorithm proposed in the present invention can realize the emission tomography three-dimensional reconstruction of the flame combustion field under the limited projection angle. The obtained reconstruction result has higher reconstruction accuracy, is superior in calculation time, and has a faster reconstruction rate.

[0015] (3) The present invention better extracts image information through the residual network, solves the problems of gradient explosion and gradient vanishing in the deep network training process, and improves the accuracy of the deep learning training model during the three-dimensional reconstruction of the combustion flame field. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a flow chart of the three-dimensional reconstruction method of the residual network finite angle combustion field based on physical constraints of the present invention.

[0017] Figure 2 This is the residual network model building diagram of the present invention, where CONV is the convolution layer, BN is BatchNorm, MAXPOOL is the maximum pooling layer, the number of input and output channels of BTNK1 is different, and the number of input and output channels of BTNK2 is the same. Compared with BTNK2, BTNK1 has an additional convolution layer on the right, which makes the input and output sizes different. Where i = 1, 2, 3, 4; when i = 1, C_1 = 8; when i = 2, C_2 = 16; when i = 3, C_3 = 32; when i = 4, C_4 = 64; S is the step size; RELU is the activation function.

[0018] Figure 3 It is the value of SSIM and RMSE of the test result and the real result when the weight coefficient λ of the regular term of the loss function takes different values ​​in the numerical simulation experiment of the algorithm of the present invention.

[0019] Figure 4This is a comparison of the reconstruction results of the algorithm of the present invention with the ART algorithm and the CNN model reconstruction algorithm under limited projection angles. Simple1, Simple2 and Simple3 are the results of the simulation field with one, two and three candles in the reconstruction area, respectively, and 3 samples are selected for a total of 9 samples.

[0020] Figure 5 This is the reprojection result of the algorithm of the present invention in the candle flame reconstruction experiment. SimpleA, simpleB and simpleC are the result analysis of one, two and three candle flames in the reconstruction area, respectively, and 7 samples are selected for a total of 21 samples.

[0021] Figure 6 The comparison of the operation time of the algorithm of the present invention and the ART algorithm is shown in Figure 1. Sample A, Sample B and Sample C are the result analysis of one, two and three candle flames in the reconstruction area, respectively, and 2 samples are selected for a total of 6 samples. DETAILED DESCRIPTION

[0022] Combination Figure 1 The present invention is a three-dimensional reconstruction method of a residual network limited angle combustion field based on physical constraints to solve the three-dimensional reconstruction problem of flame emission tomography under limited angle incomplete data projection, comprising the following steps:

[0023] Step 1: Collect N frames of continuous flame projection images through CCD cameras at M different positions of the flame emission tomography reconstruction device based on camera array. Each frame is composed of M CCD cameras collecting one flame projection image at the same time, N>500; extract the position containing the flame emission intensity feature projection in M×N flame projection images, and use the one-dimensional projection data of the same row of projection images corresponding to the position as input data. The pixel size of the input data is h×l, with a total of K images, K≥10000. Divide the input data into training set and test set in a ratio of 8:2.

[0024] Step 2: The three-dimensional structural characteristic distribution of the combustion flame field is composed of the three-dimensional numerical distribution of the emission light intensity generated by the combustion flame field. According to the three-dimensional physical structure distribution characteristics of the combustion flame field, a three-dimensional flame loss function is constructed, as follows:

[0025] Step 2-1. The chemical free radicals generated during the flame combustion process are non-negative, and the distribution of the free radicals is uniform and regular, and also has boundaries; the three-dimensional structural characteristic distribution of the combustion flame field is used as prior knowledge to constrain the flame field reconstruction result, and the prior knowledge is introduced into the loss function as a regular term.

[0026] The selected regularization term R TV Represents the sum of the gradients of all voxels in the X and Y directions of the two-dimensional distribution of the three-dimensional flame field, and the regularization term R TVThe minimization of

[0027]

[0028] Among them, u represents the two-dimensional distribution of the three-dimensional flame combustion field in the X direction and the Y direction; n represents the total number of voxels in the two-dimensional distribution of the three-dimensional flame combustion field, x represents the sequence number of the voxel in the X direction, y represents the sequence number of the voxel in the Y direction, ▽ x u is the gradient of the three-dimensional flame combustion field in the X direction, ▽ y u is the gradient of the three-dimensional flame combustion field in the Y direction.

[0029] Formula (1) is solved using the Split Bregman iterative algorithm, and the intermediate variable d x =▽ x u, intermediate variable d y =▽ y u; k represents the number of iterations. At the same time, adding a penalty function transforms equation (1) into an unconstrained optimization problem:

[0030]

[0031] Where λ1 is the relaxation factor for the control iteration, b x ,b y are all added penalty variables;

[0032] Obtain

[0033]

[0034]

[0035] Among them, the loss variable S at the kth iteration k for

[0036]

[0037] Then the updated regularization term L R for

[0038]

[0039] Step 2-2, introduce the root mean square error into the loss function:

[0040] The formula for the root mean square error is as follows

[0041]

[0042] Where T represents the size of the total data, U and P represent the size of the predicted value and the true value respectively.

[0043] Step 2-3, three-dimensional flame loss function Lloss Expressed as

[0044] L loss =L MSE +λL R (7)

[0045] Where λ represents the weight coefficient of the regularization term.

[0046] Step 3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the reconstructed three-dimensional field of flame tomography as label data into the residual network, and use the loss function to train the residual network to obtain the optimized residual network model, as follows:

[0047] Step 3-1: Use the Adam optimization algorithm to optimize the weight parameters of the residual network;

[0048] Step 3-2, set the batch size, number of batches, and learning rate;

[0049] Step 3-3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the flame tomography reconstructed three-dimensional field as label data into the residual network. According to the three-dimensional flame loss function L loss The convergence rate and function value of the residual network are gradually reduced, and the residual network is trained for multiple iterations until the residual network reaches the set number of iterations and the three-dimensional flame loss function L loss When there is no decrease for W consecutive cycles, the training ends and the residual network model is obtained, where W>50.

[0050] Step 4: Input the test set into the residual network model, and then perform three-dimensional superposition reconstruction on the two-dimensional distribution predicted by the residual network model to obtain the flame tomography three-dimensional structure distribution.

[0051] Example 1

[0052] This embodiment will illustrate the three-dimensional reconstruction method of the residual network finite angle combustion field based on physical constraints proposed in the present invention by conducting a numerical simulation reconstruction experiment on a flame model.

[0053] The device in this embodiment is composed of 12 CCD cameras arranged in an array, which can perform multi-directional detection while effectively saving the space of the reconstruction device. The 12 CCD cameras are divided into three windows, each window has four CCD cameras. The 4 CCD cameras in each window form a window in a 2x2 arrangement, with 2 CCD cameras in each column, for a total of 2 columns. The two CCD cameras in the same column are distributed in the height direction and fixed on the same fixing frame, the center of the CCD camera located below is parallel to the center of the flame, and the distance between the two CCD cameras on the same fixing frame is equal. In this embodiment, this distance is 8.5mm.

[0054] The angle between two adjacent fixing frames and the flame in the same window is 15°, and the angle between the fixing frames at the same position in two adjacent windows and the flame is 60°. The CCD camera array is centered on the flame and is distributed in an arc around the flame. The distance between each fixing frame and the flame is 600mm. The 12 CCD cameras are connected to the same computer, and the trigger program in the computer generates a pulse signal, so that the 12 CCD cameras can simultaneously collect the flame emission light intensity image at the same time.

[0055] This embodiment uses a three-dimensional flame simulation field for three-dimensional reconstruction. The three-dimensional flame simulation field is constructed by a mathematical formula. The reconstruction area where the simulation field is located is divided into 50×50×50 grids. The actual size of each grid is 0.1 mm. A total of 750 three-dimensional flame simulation fields are constructed.

[0056] Step 1: The projection images of the three-dimensional flame simulation field collected by 12 CCD cameras in three windows of the device total 12×750=9000 images; extract the position containing the projection of the flame emission intensity feature in the 9000 three-dimensional flame model projection images, and use the one-dimensional projection data of the same row of projection images corresponding to the position as the input data. The pixel size of the input data is 100×12, and there are 15000 images in total. Divide the input data into training set and test set in a ratio of 8:2;

[0057] Step 2: The three-dimensional structural characteristic distribution of the combustion flame field is composed of the three-dimensional numerical distribution of the emission light intensity generated by the combustion flame field. According to the three-dimensional physical structure distribution characteristics of the combustion flame field, a three-dimensional flame loss function is constructed.

[0058] Since the value of the weight coefficient λ of the regularization term in the loss function is different, it will affect the accuracy of the residual network training result. Therefore, in this embodiment, the weight coefficient λ of the regularization term of the loss function selects different values ​​to evaluate the loss function that can achieve the best training effect. λ is selected as 0, 0.1, 0.2…0.8 respectively for residual network training.

[0059] Step 3: Input the training set into the residual network (the residual network structure is as follows Figure 2 As shown in the figure, the two-dimensional cross section of the real distribution of the three-dimensional flame simulation field is input into the residual network as label data, and the residual network is trained using the loss function to obtain the residual network model, which is as follows:

[0060] Step 3-1: Use the Adam optimization algorithm to optimize the weight parameters of the residual network;

[0061] Step 3-2, set the batch size to 128, the number of batches to 300, and the learning rate to 0.01;

[0062] Step 3-3: Input the training set into the residual network, and input the two-dimensional cross section of the real distribution of the three-dimensional flame simulation field as label data into the residual network. According to the three-dimensional flame loss function L loss The convergence rate and function value of the residual network are set to 0.01 during the training process. After completing 500 epochs of training on all samples in the training set, the learning rate is reduced to 0.001 and the training is continued for 500 epochs. At this time, the residual network reaches the set number of iterations and the three-dimensional flame loss function L loss When there is no decrease in 100 consecutive cycles, the training ends and the optimized residual network model is obtained.

[0063] Step 4: Input the test set into the residual network model, and then perform a three-dimensional superposition on the two-dimensional distribution predicted by the residual network model to obtain the flame tomography three-dimensional structure distribution.

[0064] In order to evaluate the effect of the present invention, structural similarity (SSIM) and root mean square error (RMSE) are used as evaluation indicators. The value of SSIM is between (0, 1), and the closer it is to 1, the better the effect. The smaller the value of root mean square error (RMSE), the better the effect.

[0065] The root mean square error (RMSE) is defined as follows:

[0066]

[0067] Where N1 is the number of grids in the reconstruction area of ​​the three-dimensional flame field, j is the network number of the three-dimensional flame field, and F s is the value of the real three-dimensional flame field, F c is the value of the predicted three-dimensional flame field, (F c ) max is the maximum value in the three-dimensional distribution of the flame field. The definition of structural similarity (SSIM) is as follows:

[0068]

[0069] where μ s and μ c are the average value of the true value of the three-dimensional flame field and the average value of the predicted value of the three-dimensional flame field, σ s and σ c are the variances of the true value of the three-dimensional flame field and the predicted value of the three-dimensional flame field, σ sc is the covariance between the true value of the three-dimensional flame field and the predicted value of the three-dimensional flame field. c1 and c2 are the regularization parameters used, defined as shown in formula (10), where L represents the dynamic range of the pixel, usually 255. k1 is 0.01 and k2 is 0.03.

[0070]

[0071] A total of 30 three-dimensional flame simulation fields in the test set were selected as test samples, and the accuracy of the trained residual network model was calculated when λ took different values.

[0072] like Figure 3 As shown in the figure, when λ is 0.1, the SSIM value is the largest and the RMSE value is the smallest, which means that the residual network model has the best effect when λ=0.1.

[0073] In order to further evaluate the effect of the residual network model, we used the algebraic iteration algorithm (ART) and the convolutional neural network model (CNN) to reconstruct the three-dimensional flame simulation field of the projections collected in this embodiment and compared them with the residual network model. The three-dimensional flame loss function L in the residual network model is loss The weight coefficient of the regularization term is λ = 0.1. Figure 4 It can be seen that the algorithm of the present invention has certain advantages in reconstructing flame tomography with limited projection direction. The RMSE of the residual network model is stable at about 0.01-0.03, and the SSIM increases to about 0.9. Compared with the ART algorithm, the RMSE value is increased by about 98%, and the SSIM value is increased by about 40%. Compared with the CNN model, the RMSE and SSIM increase by about 70% and 10%, respectively. It can be seen that the residual network model has higher accuracy in the three-dimensional reconstruction of limited angle combustion fields.

[0074] Example 2

[0075] This embodiment will illustrate the method for three-dimensional reconstruction of a finite-angle combustion field using a residual network based on physical constraints proposed in the present invention through a real candle flame reconstruction experiment.

[0076] The reconstruction area where the burning candle flame is located is divided into 100×100×120 grids, and the actual size of each grid is 0.55mm.

[0077] Step 1: The device used in this embodiment is the same as the device with three windows and 12 CCD cameras in Example 1. The CCD cameras at 12 different positions in the three windows of the device collect 750 frames of continuous flame projection images, and each frame is composed of 12 CCD cameras collecting one flame projection image at the same time; extract the position containing the flame emission intensity feature projection in 9,000 flame projection images, and use the one-dimensional projection data of the same row of projection images corresponding to the position as input data. The pixel size of the input data is 150×12, and there are 15,000 images in total. Divide the input data into a training set and a test set in a ratio of 8:2;

[0078] Step 2: The three-dimensional structural characteristic distribution of the combustion flame field is composed of the three-dimensional numerical distribution of the emission light intensity generated by the combustion flame field. According to the three-dimensional physical structure distribution characteristics of the combustion flame field, a three-dimensional flame loss function is constructed.

[0079] Three-dimensional flame loss function L loss Expressed as

[0080] L loss =L MSE +0.1L R

[0081] Step 3: The device for obtaining the real distribution of the candle flame tomography reconstruction three-dimensional field is composed of 12 CCD cameras evenly arranged within a range of 180°. The 12 CCD cameras are fixed on 12 fixing frames, and the center of each CCD camera is parallel to the center of the flame. The angle between the two adjacent fixing frames and the flame is 15°. The 12 fixing frames are distributed in an arc around the flame, and the distance between each fixing frame and the flame is 600mm.

[0082] The CCD camera in this device and the three window CCD cameras in the above device are connected to the same computer. The trigger program in the computer generates a pulse signal, which enables multiple CCD cameras to simultaneously collect flame emission light intensity images at the same time. The ART algorithm is the most commonly used reconstruction method in emission spectral tomography technology. After obtaining multi-directional and multi-angle projections of the combustion field, it can accurately reconstruct the three-dimensional structure distribution of the combustion field.

[0083] The candle flame projections collected by 12 evenly distributed CCD cameras within a 180° range are reconstructed in three dimensions using the ART algorithm. The resulting three-dimensional structure distribution of the candle flame is the true distribution of the three-dimensional field reconstructed by the candle flame tomography.

[0084] Input the training set into the residual network (the residual network structure is as follows Figure 2 As shown in Figure 2, the two-dimensional cross section of the true distribution of the reconstructed three-dimensional field of flame tomography is input into the residual network as label data, and the residual network is trained using the loss function to obtain an optimized residual network model, as follows:

[0085] Step 3-1: Use the Adam optimization algorithm to optimize the weight parameters of the residual network;

[0086] Step 3-2, set the batch size to 128, the number of batches to 500, and the learning rate to 0.01;

[0087] Step 3-3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the flame tomography reconstructed three-dimensional field as label data into the residual network. According to the three-dimensional flame loss function L lossThe convergence rate and function value of the residual network are set to 0.01 during the training process. After completing 500 epochs of training on all samples in the training set, the learning rate is reduced to 0.001 and the training is continued for 500 epochs. At this time, the residual network reaches the set number of iterations and the three-dimensional flame loss function L loss When there is no decrease in 100 consecutive cycles, the training ends and the residual network model is obtained.

[0088] Step 4: Input the test set into the residual network model, and then perform three-dimensional superposition reconstruction on the two-dimensional distribution predicted by the residual network model to obtain the flame tomography three-dimensional structure distribution.

[0089] The candle flame tomography three-dimensional distribution reconstructed by the algorithm of the present invention is projected onto the target surface of one of the cameras, and the structural similarity (SSIM) value between the projection image and the projection image collected by the camera is obtained for quantitative evaluation. Figure 5 As shown, the obtained SSIM values ​​are all greater than 0.97, indicating that the algorithm of the present invention has a high reconstruction accuracy.

[0090] Figure 6 The results of the algorithm of the present invention and the ART algorithm are compared. We can see that the residual network model has outstanding advantages in computational efficiency compared with the traditional ART reconstruction algorithm. The computation time used to reconstruct the three-dimensional distribution of the candle flame is increased to about 0.1 seconds, which is less than 1% of the computation time of the traditional ART algorithm.

Claims

1. A three-dimensional reconstruction method of finite angle combustion field based on residual network with physical constraints, characterized in that: The following steps are involved: Step 1: Capture N frames of continuous flame projection images through CCD cameras at M different positions of a flame emission tomography reconstruction device based on a camera array, where each frame is composed of M CCD cameras simultaneously capturing one flame projection image, N>500; extract the position containing the flame emission intensity projection in the M×N flame projection images, and use the one-dimensional projection data of the same row of projection images corresponding to the position as input data, where the pixel size of the input data is h×l, with a total of K images, K≥10000; divide the input data into a training set and a test set in a ratio of 8:2; Step 2: According to the three-dimensional physical structure distribution characteristics of the combustion flame field, a three-dimensional flame loss function is constructed; the details are as follows: Step 2-1, the chemical free radicals generated during the flame combustion process are non-negative, and the distribution of the free radicals is uniform and regular, and also has boundaries; the three-dimensional structural characteristic distribution of the combustion flame field is used as the prior knowledge constraining the flame field reconstruction result, and the prior knowledge is introduced into the loss function as a regular term; The selected regularization term R TV Represents the sum of the gradients of all voxels in the X and Y directions of the two-dimensional distribution of the three-dimensional flame field, and the regularization term R TV The minimization of Among them, u represents the two-dimensional distribution of the three-dimensional flame combustion field in the X direction and the Y direction; n represents the total number of voxels in the two-dimensional distribution of the three-dimensional flame combustion field, x represents the sequence number of the voxel in the X direction, and y represents the sequence number of the voxel in the Y direction. is the gradient of the three-dimensional flame combustion field in the X direction, is the gradient of the three-dimensional flame combustion field in the Y direction; Formula (1) is solved using the Split Bregman iterative algorithm, with the intermediate variable Intermediate variables At the same time, adding a penalty function transforms equation (1) into an unconstrained optimization problem: Where λ1 is the relaxation factor for the control iteration, b x ,b y are all added penalty variables; Obtain Among them, the superposition variable S at the kth iteration k for Then the updated regularization term L R for Step 2-2, introduce the root mean square error into the loss function: Root mean square error L MSE The formula is as follows Among them, T represents the size of the total amount of data, U represents the predicted value, and P represents the true value; Step 2-3, three-dimensional flame loss function L loss Expressed as THE loss =L MSE +λL R (7) Where λ represents the weight coefficient of the regularization term; Step 3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the reconstructed three-dimensional field of flame tomography as label data into the residual network, and use the loss function to train the residual network to obtain an optimized residual network model; the details are as follows: Step 3-1: Use the Adam optimization algorithm to optimize the weight parameters of the residual network; Step 3-2, set the batch size, number of batches, and learning rate; Step 3-3: Input the training set into the residual network, and input the two-dimensional cross section of the true distribution of the flame tomography reconstructed three-dimensional field as label data into the residual network. According to the three-dimensional flame loss function L loss The convergence rate and function value of the residual network are gradually reduced, and the residual network is trained for multiple iterations until the residual network reaches the set number of iterations and the three-dimensional flame loss function L loss When there is no decrease for W consecutive cycles, the training ends and the residual network model is obtained, W>50; Step 4: Input the test set into the residual network model, and then perform three-dimensional superposition reconstruction on the two-dimensional distribution predicted by the residual network model to obtain the flame tomography three-dimensional structure distribution.

2. The method for three-dimensional reconstruction of a combustion field with a residual network based on physical constraints according to claim 1 is characterized in that: In step 2, the three-dimensional structural characteristic distribution of the combustion flame field is composed of the three-dimensional numerical distribution of the emission light intensity generated by the combustion flame field.

Images