Blast furnace carbon emission concentration prediction method based on digital twinning

Abstract

The invention provides a blast furnace carbon emission concentration prediction method based on digital twinning, and relates to the technical field of digital twinning. Multi-source data fusion is adopted, a simple mechanism model is used for building a blast furnace temperature field, and the blast furnace carbon emission concentration prediction method based on digital twinning is obtained by combining operation parameters with image key features of a blast furnace top thermal imaging image and supplementing temperature field features. Establishing a shared matrix; and extracting image features by using ResNet50 and an auto-encoder, and establishing a random forest (RF) model. Based on a shared matrix, a random weight neural network algorithm (RVFLNs) model is constructed, and prediction results of a random forest (RF) model and the random weight neural network algorithm (RVFLNs) model are combined by adopting a particle swarm optimization (PSO) algorithm, so that the prediction precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twins, and in particular to a method for predicting carbon emission concentration of a blast furnace based on digital twins. Background Art

[0002] The blast furnace is one of the most critical equipment in the steel production process. Its operating efficiency and emission control are crucial to increasing steel production, saving energy, reducing emissions and lowering production costs. With the introduction of the country's dual carbon goals, accurately predicting the CO2 and CO emission concentrations of blast furnaces is crucial to the sustainable development of the steel industry and achieving carbon neutrality goals. Researchers have proposed various methods for predicting blast furnace carbon emissions, which can be divided into two categories: mechanism models and data-driven models. The former relies on the understanding and simulation of the physical and chemical processes inside the blast furnace, while the latter uses statistical and machine learning techniques to analyze and predict emission patterns based on historical data.

[0003] Mechanistic modeling approaches involve an in-depth understanding and simulation of the complex physical and chemical processes within a blast furnace, and can effectively predict the concentrations of CO2 and CO in blast furnace emissions. Wu et al. used a variable activation energy model (VAEM) to experimentally study the molten reduction reaction of coal coke and iron ore under simulated blast furnace conditions. By analyzing the kinetic behavior of the molten reduction process, they predicted the carbon oxides in blast furnace exhaust gas. Xia et al. established a mathematical model including the blast furnace shaft, combustion zone, and gas circulation system to estimate the composition of blast furnace emissions by analyzing the heat transfer, combustion reaction, and gas circulation in the blast furnace. Li et al. 6 adopted the all-oxygen blast furnace technology and developed a carbon metabolism and prediction model to study and analyze the changes in CO2 and CO in blast furnace emissions and their impact on carbon emissions. Jiang et al. 7 constructed a mathematical model of mass and energy balance of the blast furnace and a heat balance model of the hot blast stove. They used metallurgical thermodynamics methods to predict and analyze the composition of blast furnace gas.

[0004] Although mechanistic modeling can effectively explain the production process of carbon oxide, it also has some disadvantages. First, mechanistic models are poorly adaptable to changes in blast furnace operating conditions and are difficult to update in real time. Second, these models require a lot of computing resources, making them difficult to run in large-scale industrial applications.

[0005] In recent years, there has been an increasing number of studies on the use of advanced data-driven models to predict blast furnace gas composition. These methods can provide more accurate and real-time predictions through in-depth analysis of historical data and operating parameters, thereby optimizing blast furnace operations. Zhang et al. proposed a temporal dual graph convolutional network (TDGCN) model combining graph convolution, hypergraph convolution, and TimesNet to predict CO2 and CO content in blast furnace gas. Xiao et al. developed a chaotic radial basis function (RBF) model for predicting carbon monoxide utilization rate (CMUR). By reconstructing the phase space of CMUR, they used chaos theory to achieve high-precision predictions. Sha et al. introduced a prediction method based on kernel principal component analysis (KPCA) and extreme learning machine (ELM) for carbon monoxide utilization. They selected important input variables through KPCA and developed a prediction model using ELM. Feng et al.11 proposed a data-driven model for blast furnace gas production prediction based on spectral decomposition, which significantly improved the prediction accuracy through spectral decomposition and combined modeling. Yang et al. 12 adopted a multi-factor prediction model based on generalized phase permutation entropy (GPPE) and singular spectrum decomposition (SSD), combined with a bidirectional long short-term memory network (GPPE-SSD-TDO-BiLSTM) optimized using an artificial Tasmanian devil optimizer to predict the CO2 and CO contents in blast furnace gas.

[0006] Although these methods have achieved significant progress in prediction accuracy and real-time performance, they fail to fully utilize the collaborative modeling capabilities of blast furnace top-view thermal imaging. Top-view thermal imaging not only provides detailed information on the temperature distribution inside the blast furnace, but also indirectly reflects the combustion and emission patterns in different areas during operation. Ignoring this key data source may lead to defects in the prediction model in dealing with carbon emission fluctuations and optimizing control strategies. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a method for predicting carbon emission concentration of a blast furnace based on digital twins, comprising:

[0008] Step 1: Real-time collection of production data during blast furnace smelting, wherein the production data includes operating parameters, tuyere parameters and multiple blast furnace top thermal images, wherein the tuyere parameters include tuyere temperature and tuyere velocity;

[0009] Step 2: Screen all the thermal images of the blast furnace top to obtain the thermal images that are not blocked by the charging chute;

[0010] Step 3: According to the thermal imaging image and tuyere parameters not blocked by the charging chute, the temperature field characteristics are obtained, wherein the temperature field characteristics include temperature gradient, maximum temperature and average temperature;

[0011] Step 4: For each thermal image that is not blocked by the charging chute, extract the features through ResNet50 and autoencoder to obtain the image features of all thermal images that are not blocked by the charging chute, and compress them into an image feature matrix X. The independent component analysis algorithm is used to extract the image feature matrix X to obtain the non-Gaussian independent source feature S. The non-Gaussian independent source feature S is processed by ICA inverse transformation to obtain the image key feature X. e , the key feature X of the image e Combined with the physical variable Z, we get the shared matrix X c =[Z;X e ], where the physical variable Z includes the operating parameters and temperature field characteristics after preprocessing, and the principal component analysis algorithm is used to extract the Gaussian features in the shared matrix;

[0012] Step 5: Obtain the final carbon monoxide concentration and carbon dioxide concentration based on the non-Gaussian independent source feature S and the Gaussian feature.

[0013] Optionally, step 2 specifically includes:

[0014] Through the gray level co-occurrence matrix, feature extraction is performed on each of the blast furnace top thermal images to obtain the texture features of each blast furnace top thermal image, wherein the texture features include contrast, energy, uniformity and the difference of pixels in the blast furnace top thermal image. According to the blast furnace top thermal image and the corresponding texture features, all the blast furnace top thermal images are screened to obtain the thermal images that are not blocked by the charging chute.

[0015] Optionally, all thermal images of the blast furnace top are screened according to the thermal images of the blast furnace top and the corresponding texture features to obtain thermal images that are not blocked by the charging chute, including:

[0016] According to the thermal imaging image of the blast furnace top, the production stage of the blast furnace is determined. When the blast furnace is in the unloading stage, the value range of texture features not blocked by the charging chute corresponding to the unloading stage is obtained, and it is used as the target value range. When the blast furnace is in the charging stage, the value range of texture features not blocked by the charging chute corresponding to the charging stage is obtained, and it is used as the target value range; obtain the thermal imaging image of the blast furnace top with texture features within the target value range, and obtain the thermal imaging image not blocked by the charging chute.

[0017] Optionally, step 3 specifically includes:

[0018] The mass conservation equation, momentum conservation equation and energy conservation equation are used to process the tuyere parameters to obtain the initial temperature field. For the flame area in the thermal imaging image that is not blocked by the charging chute, the grayscale value of the flame area is extracted, and the actual temperature value corresponding to the grayscale value is obtained. The actual temperature value is combined with the initial temperature field to obtain the final temperature field. Feature extraction is performed on the final temperature field to obtain the temperature field characteristics.

[0019] Optionally, step 5 specifically includes:

[0020] Step 5.1: Process the non-Gaussian independent source feature S in step 4 by using a first prediction model based on a random forest regression algorithm to obtain a first concentration prediction value of carbon monoxide and a first concentration prediction value of carbon dioxide;

[0021] The first prediction model is obtained by model training based on multiple first training samples, the first training samples include a first input sample and a first output sample, the first input sample is a historical non-Gaussian independent source feature, and the first output sample includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration corresponding to the historical non-Gaussian independent source feature;

[0022] Step 5.2: The Gaussian features in step 4 are processed by a plurality of different second prediction models based on a random weight neural network algorithm to obtain a plurality of second prediction values, wherein the second prediction values ​​include a second concentration prediction value of carbon monoxide and a second concentration prediction value of carbon dioxide. According to a preset weight, all the second concentration prediction values ​​of carbon monoxide are weighted averaged to obtain a third concentration prediction value of carbon monoxide, and all the second concentration prediction values ​​of carbon dioxide are weighted averaged to obtain a third concentration prediction value of carbon dioxide.

[0023] Step 5.3: Determine the first weight and the second weight through the particle swarm optimization algorithm PSO, and perform weighted summation on the first predicted value of carbon monoxide concentration and the third predicted value of carbon monoxide concentration according to the first weight and the second weight to obtain the final carbon monoxide concentration, and perform weighted summation on the first predicted value of carbon dioxide concentration and the third predicted value of carbon dioxide concentration according to the first weight and the second weight to obtain the final carbon dioxide concentration.

[0024] Optionally, in step 5.3, a plurality of different second prediction models are obtained by the following steps:

[0025] Step A1: obtaining a historical Gaussian feature and a true concentration value corresponding to the historical Gaussian feature, wherein the true concentration value includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration, taking the historical Gaussian feature as a second input sample, taking the true concentration value as a second output sample, the second input sample and the second output sample constitute a second training sample, and then obtaining a plurality of second training samples;

[0026] Step A2: adopt a random sampling strategy with replacement to perform bagging sampling on multiple second training samples to obtain multiple second training samples to form a sub-training set, and then obtain multiple sub-training sets;

[0027] Step A3: for each sub-training set, construct an initial model of a random weight neural network algorithm, wherein the initial model weights and biases are random;

[0028] Step A4: The initial model is trained through a sub-training set. Specifically, the input samples in the sub-training set are input into the initial model to obtain historical prediction values ​​corresponding to historical Gaussian features, wherein the historical prediction values ​​include predicted values ​​of carbon monoxide concentration and predicted values ​​of carbon dioxide concentration. According to the historical prediction values ​​and the output samples, a regularized least squares solution algorithm is used to update the parameters in the initial model until the model converges to obtain a second prediction model, and then multiple different second prediction models corresponding to multiple sub-training sets are obtained.

[0029] Optionally, in step 5.3, determining the first weight and the second weight by using a particle swarm optimization algorithm PSO includes:

[0030] Step B1: obtaining a predicted value of carbon monoxide concentration and a true value of carbon monoxide concentration, as well as a predicted value of carbon dioxide concentration and a true value of carbon dioxide concentration, wherein the predicted value of carbon monoxide concentration includes a first concentration value and a second concentration value, and the predicted value of carbon dioxide includes a third concentration value and a fourth concentration value;

[0031] Step B2: Initialize the speed and position of all particles in the particle swarm. The position of each particle includes two randomly initialized weights. The individual optimal position and the global optimal position are initialized, and the initial number of iterations is set.

[0032] Step B3: the initial speed of the particle is used as the current speed, the initial position of the particle is used as the current position, the initial individual optimal position of each particle is used as the current individual optimal position of the particle, the initial global optimal position is used as the current global optimal position, and the initial number of iterations is used as the current number of iterations;

[0033] Step B4: for each particle, according to the weight in the current position of the particle, weighted sum of the first concentration value and the second concentration value is performed to obtain a weighted predicted value of the carbon monoxide concentration, and weighted sum of the third concentration value and the fourth concentration value is performed to obtain a weighted predicted value of the carbon dioxide concentration;

[0034] Step B5: for each particle, the mean square error between the weighted predicted value of carbon monoxide concentration and the true value of carbon monoxide concentration is calculated to obtain a first error, and the mean square error between the weighted predicted value of carbon dioxide concentration and the true value of carbon dioxide concentration is calculated to obtain a second error. The first error and the second error constitute the current fitness of the particle. Among the current fitness of the particle and the fitness corresponding to the current individual optimal position of the particle, the position with the smaller fitness is selected as the new current individual optimal position of the particle.

[0035] Step B6: Among the fitness of the current individual optimal position of all particles after update and the fitness of the current global optimal position, select the position with the smaller fitness as the new current global optimal position;

[0036] Step B7: Update the current position and current speed of each particle, which is specifically achieved through the following formula:

[0037] v i,t+1 =ωv i,t +c1r1(p best,i -x i,t )+c2r2(g best -x i,t )

[0038] x i,t+1 =x i,t +v i,t+1

[0039] Among them, ω is the inertia weight, c1, c2 are acceleration constants, r1, r2 are random numbers, v i,t+1 is the current velocity of the ith particle at iteration t+1, v i,t is the current velocity of the ith particle at the current iteration number t, p best,i is the current individual optimal position of the i-th particle at the current iteration number t, g best is the current global optimal position of the ith particle at the current iteration number t, x i,t is the current position of the ith particle at the current iteration number t, x i,t+1 is the current position of the i-th particle at iteration t+1;

[0040] Step B8: The current number of iterations is increased by one. When the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold, the two weights in the new current global optimal position are used as the first weight and the second weight. When the current number of iterations is not less than the preset number, and the fitness of the new current global optimal position is not less than the preset threshold, return to execute step B4 until the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold.

[0041] Optionally, an independent component analysis algorithm is used to extract the image feature matrix X to obtain non-Gaussian independent source features S, which is specifically implemented by the following formula:

[0042] S = WX;

[0043] Where W is the unmixing matrix.

[0044] Optionally, the ICA inverse transform is used to process the non-Gaussian independent source feature S to obtain the image key feature X e , which is specifically achieved through the following formula:

[0045] X e＝ X-AS = (I-AW)X;

[0046] Among them, A is the mixing coefficient matrix and I is the unit matrix.

[0047] The beneficial effects of adopting the above technical solution are:

[0048] The present invention proposes a method for predicting carbon emission concentration of blast furnace based on digital twin. The present invention adopts multi-source data fusion and uses a simple mechanism model to construct the temperature field of blast furnace. By combining the operating parameters with the key features of the image of the thermal imaging image of the blast furnace top and supplementing the temperature field features, a shared matrix is ​​established; then ResNet50 and autoencoder are used to extract image features and a random forest (RF) model is established. Based on the shared matrix, a random weighted neural network algorithm (RVFLNs) model is constructed, and the prediction results of the random forest (RF) model and the random weighted neural network algorithm (RVFLNs) model are combined by the particle swarm optimization (PSO) algorithm to improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic flow chart of a method for predicting carbon emission concentration of a blast furnace based on digital twin in an embodiment of the present invention;

[0050] Figure 2: are thermal images of the unloaded furnace in an embodiment of the present invention, wherein (a) is a thermal image of the blast furnace top in the 1st frame when unloaded, (b) is a thermal image of the blast furnace top in the 50th frame when unloaded, (c) is a thermal image of the blast furnace top in the 60th frame when unloaded, (d) is a thermal image of the blast furnace top in the 70th frame when unloaded, (e) is a thermal image of the blast furnace top in the 80th frame when unloaded, and (f) is a thermal image of the blast furnace top in the 100th frame when unloaded;

[0051] Figure 3 : are thermal imaging images during charging in an embodiment of the present invention, wherein (a) is a thermal imaging image of the blast furnace top at the 1st frame during charging, (b) is a thermal imaging image of the blast furnace top at the 50th frame during charging, (c) is a thermal imaging image of the blast furnace top at the 60th frame during charging, (d) is a thermal imaging image of the blast furnace top at the 70th frame during charging, (e) is a thermal imaging image of the blast furnace top at the 80th frame during charging, and (f) is a thermal imaging image of the blast furnace top at the 100th frame during charging;

[0052] Figure 4 is a texture feature map of the unloaded stage in an embodiment of the present invention;

[0053] Figure 5 This is a representative sample diagram randomly selected during the unloaded stage in the embodiment of the present invention;

[0054] Figure 6 is a texture feature map of the loading stage in an embodiment of the present invention;

[0055] Figure 7 It is a representative sample diagram randomly selected during the charging stage in the embodiment of the present invention;

[0056] Figure 8 is a schematic diagram of the final temperature field in an embodiment of the present invention;

[0057] Fig. 9 is a flow chart of a shared ICA-PCA algorithm in an embodiment of the present invention;

[0058] Fig.10 is a diagram showing the proportion of carbon dioxide in the gas composition in an embodiment of the present invention;

[0059] Fig.11 is a diagram showing the proportion of carbon monoxide in the gas composition in an embodiment of the present invention;

[0060] Fig.12 is a distribution diagram of carbon dioxide (CO2) prediction errors in an embodiment of the present invention;

[0061] Fig.13 4 is a carbon dioxide (CO) prediction error distribution diagram in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0063] The blast furnace ironmaking process is crucial in steel production. The blast furnace has a complex internal structure and operates at high temperature and high pressure, with strong coupling and significant hysteresis. It involves multiple factors that directly or indirectly affect carbon emissions, such as temperature gradient, material flow, and thermodynamic equilibrium. These factors interact with each other and directly affect the formation and distribution of carbon emissions. Therefore, before establishing a carbon emission prediction model, it is necessary to have a deep understanding of the internal structure and production process of the blast furnace in order to establish an accurate prediction model and achieve precise control.

[0064] The main components of blast furnace emissions include top gas components such as hydrogen, nitrogen, carbon monoxide and carbon dioxide. Hydrogen mainly acts as a reducing agent in the reaction, but the concentration is relatively low. Nitrogen mainly plays a diluting role and has the least direct impact on the blast furnace reaction process. Therefore, the focus is more on the concentration of carbon monoxide and carbon dioxide and their impact on blast furnace operation.

[0065] Carbon monoxide is a key reductant in the iron ore reduction process in the blast furnace, which can improve the reduction rate of iron ore and increase the output of molten iron. However, too high CO concentrations may indicate incomplete combustion, leading to increased coke consumption and potential safety hazards. CO2 is the main product of coke combustion, and its concentration reflects the combustion conditions. High concentrations of CO2 indicate complete combustion and improved thermal efficiency, but too high CO2 levels may lead to uneven heat distribution, affecting thermodynamic equilibrium and the reduction process of iron ore in the blast furnace.

[0066] The structure of the blast furnace, the generation and treatment of monoxide and dioxide are as follows:

[0067] 1. Feeding system: Raw materials such as ore and coke are fed into the blast furnace through the feeding system, ensuring a continuous supply of materials.

[0068] 2. Top feeding: The material enters the top of the furnace and gradually descends due to gravity.

[0069] 3. Blast furnace body:

[0070] Throat and shaft: In the throat and shaft, the raw material begins to heat. During this preheating stage, no significant chemical reaction occurs.

[0071] Abdomen and waist: As the material descends further, the temperature rises, the coke begins to burn, producing carbon dioxide and releasing a lot of heat. The chemical reactions here are mainly:

[0072] C+O2→CO2;

[0073] 4. Reduction reaction: As the temperature further decreases and the carbon dioxide reacts with the coke to produce carbon monoxide:

[0074] CO2+2C→2CO;

[0075] This endothermic reaction occurs in the high temperature area, especially in the lower part and waist of the furnace.

[0076] 5. Hot air system: Hot air is blown into the bottom of the blast furnace through the tuyere, and burns violently with the descending coke, producing a large amount of carbon monoxide and carbon dioxide, thereby increasing the temperature inside the furnace.

[0077] 6. Reaction process: Carbon monoxide reacts with iron ore (iron oxide) in the furnace, reducing it to iron and producing CO:

[0078] 3CO+Fe2O3→2Fe+3CO2;

[0079] The generated carbon dioxide continues to react with coke to generate more carbon dioxide.

[0080] 7. Blast furnace gas treatment system and gas recovery device: The blast furnace gas treatment system treats and recovers the gas in the furnace through gravity dust removal, silencer and water seal device. The treated blast furnace gas can be used for power generation or as fuel.

[0081] Accurately predicting the concentration of CO and CO2 in blast furnace emissions is crucial to the safety, environmental protection and economic efficiency of blast furnace operations. By real-time monitoring and predicting the concentration of these gases, operators can adjust operating parameters in a timely manner to achieve optimal control, reduce coke consumption, reduce production costs, effectively reduce CO2 emissions, avoid safety risks, and ensure the safe operation of blast furnaces.

[0082] Based on this, the present invention provides a method for predicting carbon emission concentration of blast furnace based on digital twin. Figure 1 , which may include the following steps:

[0083] Step 1: Real-time collection of production data during blast furnace smelting, the production data including operating parameters, tuyere parameters and multiple blast furnace top thermal images, the tuyere parameters including tuyere temperature and tuyere speed; the production data is Figure 1 Actual production data of medium and high-blast furnace smelting;

[0084] Among them, the operating parameters include 24 parameters such as oxygen concentration, central feed amount, blowing intensity, and furnace static pressure.

[0085] The experimental data of the present invention comes from the actual production data of No. 7 blast furnace of a domestic steel plant. A flame video acquisition module is installed inside the blast furnace to obtain the thermal imaging image of the blast furnace top during the actual production process. The video frame rate is 12.5 frames per second, and the resolution of each frame is 384×288RGB three-channel color image. While collecting the video, the computer records and saves the corresponding blast furnace production and operation parameters into the database.

[0086] The collected thermal images of the blast furnace top show that the flame is blocked by the charging chute of the blast furnace. Figure 2 Figure (c) in Figure 3 Figure (c) in which, Figure 2 are thermal imaging images without loading, wherein, Figure (a) is a thermal imaging image of the blast furnace top in the 1st frame without loading, Figure (b) is a thermal imaging image of the blast furnace top in the 50th frame without loading, Figure (c) is a thermal imaging image of the blast furnace top in the 60th frame without loading, Figure (d) is a thermal imaging image of the blast furnace top in the 70th frame without loading, Figure (e) is a thermal imaging image of the blast furnace top in the 80th frame without loading, and Figure (f) is a thermal imaging image of the blast furnace top in the 100th frame without loading; Figure 3 are thermal images during charging, wherein, Figure (a) is the thermal image of the blast furnace top in the 1st frame during charging, Figure (b) is the thermal image of the blast furnace top in the 50th frame during charging, Figure (c) is the thermal image of the blast furnace top in the 60th frame during charging, Figure (d) is the thermal image of the blast furnace top in the 70th frame during charging, Figure (e) is the thermal image of the blast furnace top in the 80th frame during charging, and Figure (f) is the thermal image of the blast furnace top in the 100th frame during charging.

[0087] The flame in the thermal image is blocked by the charging chute of the blast furnace, which will reduce the accuracy of subsequent processing. It is necessary to screen the thermal image of the blast furnace top, which is specifically achieved through step 2.

[0088] Step 2: Screen all the thermal images of the blast furnace top to obtain the thermal images that are not blocked by the charging chute;

[0089] Through the gray level co-occurrence matrix, feature extraction is performed on each of the blast furnace top thermal images to obtain the texture features of each blast furnace top thermal image, wherein the texture features include contrast, energy, uniformity and the difference of pixels in the blast furnace top thermal image. According to the blast furnace top thermal image and the corresponding texture features, all the blast furnace top thermal images are screened to obtain the thermal images that are not blocked by the charging chute.

[0090] Among them, according to the blast furnace top thermal image and the corresponding texture features, all blast furnace top thermal images are screened to obtain thermal images that are not blocked by the charging chute, including:

[0091] According to the thermal imaging image of the blast furnace top, the production stage of the blast furnace is determined. When the blast furnace is in the unloading stage, the value range of texture features not blocked by the charging chute corresponding to the unloading stage is obtained, and it is used as the target value range. When the blast furnace is in the charging stage, the value range of texture features not blocked by the charging chute corresponding to the charging stage is obtained, and it is used as the target value range; obtain the thermal imaging image of the blast furnace top with texture features within the target value range, and obtain the thermal imaging image not blocked by the charging chute.

[0092] In the specific implementation process, the texture characteristics of the unloaded stage are as follows: Figure 4 , among which, there are 7 images that meet the requirements, and the representative sample randomly selected is the 34th frame, as shown in Figure 5 The texture characteristics of the loading stage are shown in Figure 6 , among which, there are 5 images that meet the requirements. A representative sample randomly selected, namely the 83rd frame, is shown in Figure 7 shown.

[0093] Step 3: According to the thermal imaging image and tuyere parameters not blocked by the charging chute, the temperature field characteristics are obtained, wherein the temperature field characteristics include temperature gradient, maximum temperature and average temperature;

[0094] Wherein, step 3 specifically includes:

[0095] The mass conservation equation, momentum conservation equation and energy conservation equation are used to process the tuyere parameters to obtain the initial temperature field. For the flame area in the thermal imaging image that is not blocked by the charging chute, the grayscale value of the flame area is extracted, and the actual temperature value corresponding to the grayscale value is obtained. The actual temperature value is combined with the initial temperature field to obtain the final temperature field, such as Figure 8 , extract the features of the final temperature field and obtain the temperature field features.

[0096] In specific implementation, the energy conservation equation needs to be simplified to obtain a simplified energy conservation equation, which can be specifically expressed by the following formula:

[0097]

[0098] The mass conservation equation is expressed by the following formula:

[0099]

[0100] The momentum conservation equation is expressed by the following formula:

[0101]

[0102] Step 4: Combine Fig. 9, each thermal image that is not blocked by the charging chute is extracted through ResNet50 and autoencoder to obtain the image features of all thermal images that are not blocked by the charging chute, and then compressed into an image feature matrix X. The image feature matrix X is extracted using the independent component analysis algorithm to obtain the non-Gaussian independent source feature S. The non-Gaussian independent source feature S is processed using the inverse ICA transform to obtain the image key feature X e , the key feature X of the image e Combined with the physical variable Z, we get the shared matrix X c =[Z;X e ], where the physical variable Z includes the operating parameters and temperature field characteristics after preprocessing, and the principal component analysis algorithm is used to extract the Gaussian features in the shared matrix; where the operating parameters are Figure 1 After data preprocessing, the operating parameters in the data are obtained, such as oxygen enrichment rate, central ore addition amount, blast kinetic energy and so on, which are the operating parameters after preprocessing.

[0103] Among them, the independent component analysis algorithm is used to extract the image feature matrix X to obtain the non-Gaussian independent source feature S, which is specifically implemented by the following formula:

[0104] S = WX;

[0105] Where W is the unmixing matrix.

[0106] Among them, ICA inverse transform is used to process the non-Gaussian independent source feature S to obtain the key feature X of the image e , which is specifically achieved through the following formula:

[0107] X e＝ X-AS = (I-AW)X;

[0108] Among them, A is the mixing coefficient matrix and I is the unit matrix.

[0109] In the specific implementation process, the cumulative variance contribution rate of the principal component is shown in Table 1:

[0110] Table 1 Cumulative variance contribution of principal components

[0111] Main Ingredients Variance contribution rate (%) PCA 1 33.38 PCA 2 29.41 PCA 3 10.90 PCA4 6.16 PCA 5 6.02 PCA6 4.66 PCA 7 3.49 PCA 8 1.81 Cumulative variance contribution (%) 95.82

[0112] The Gaussian features obtained by dimensionality reduction of multi-source data through principal component analysis (PCA) and the non-Gaussian features extracted from image data through independent component analysis (ICA) are used as inputs to the prediction model. In order to improve the prediction accuracy, an optimization algorithm is used to determine the weight of each model, thereby achieving an accurate prediction of the ratio of CO2 and CO in blast furnace carbon emissions.

[0113] Step 5: Obtain the final carbon monoxide concentration and carbon dioxide concentration based on the non-Gaussian independent source feature S and the Gaussian feature.

[0114] Step 5 specifically includes:

[0115] Step 5.1: Process the non-Gaussian independent source feature S in step 4 by using a first prediction model based on a random forest regression algorithm to obtain a first concentration prediction value of carbon monoxide and a first concentration prediction value of carbon dioxide;

[0116] The first prediction model is obtained by model training based on multiple first training samples, the first training samples include a first input sample and a first output sample, the first input sample is a historical non-Gaussian independent source feature, and the first output sample includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration corresponding to the historical non-Gaussian independent source feature;

[0117] Random Forest (RF) is an ensemble learning algorithm, specifically referred to as "Random Forests", which uses multiple decision trees for prediction and aggregates their results to obtain the final prediction result. This method solves the overfitting problem associated with a single decision tree, thereby improving the accuracy and generalization ability of the model. Random Forest performs well in processing high-dimensional and complex data, and is very suitable for predicting blast furnace carbon emission concentration.

[0118] Traditional neural network models perform well in dealing with complex and highly uncertain objects, but training is usually time-consuming and prone to problems such as local optimality, overfitting, and poor robustness. However, the carbon emission prediction model requires real-time prediction results. In order to solve these problems, Bao et al. proposed a random weight neural network algorithm (RVFLNs) structure. RVFLNs is a feedforward neural network with only one hidden layer. The weights and biases from the input layer to the hidden layer are randomly generated within a certain range, while the weights from the hidden layer to the output layer are solved by the least squares method, avoiding the iterative process. Therefore, RVFLNs has the characteristics of short training time and strong generalization ability, which is specifically achieved through step 5.2.

[0119] Step 5.2: The Gaussian features in step 4 are processed by multiple different second prediction models based on a random weight neural network algorithm to obtain multiple second prediction values, wherein the second prediction values ​​include a second concentration prediction value of carbon monoxide and a second concentration prediction value of carbon dioxide. According to preset weights, all second concentration prediction values ​​of carbon monoxide are weighted averaged to obtain a third concentration prediction value of carbon monoxide, and all second concentration prediction values ​​of carbon dioxide are weighted averaged to obtain a third concentration prediction value of carbon dioxide. The specific reference for the random weight neural network algorithm is "Functional-LinkNet Computing: Theory, System Architecture, and Functionalities".

[0120] Step 5.3: Determine the first weight and the second weight by using the particle swarm optimization algorithm PSO. According to the first weight and the second weight, perform weighted summation on the first predicted value of carbon monoxide concentration and the third predicted value of carbon monoxide concentration to obtain the final carbon monoxide concentration. According to the first weight and the second weight, perform weighted summation on the first predicted value of carbon dioxide concentration and the third predicted value of carbon dioxide concentration to obtain the final carbon dioxide concentration. The final carbon monoxide concentration is Figure 1 The CO content in the blast furnace exhaust gas, the final carbon dioxide concentration is Figure 1 CO2 content in blast furnace exhaust gas.

[0121] In step 5.3, a plurality of different second prediction models are obtained by the following steps:

[0122] Step A1: obtaining a historical Gaussian feature and a true concentration value corresponding to the historical Gaussian feature, wherein the true concentration value includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration, taking the historical Gaussian feature as a second input sample, taking the true concentration value as a second output sample, the second input sample and the second output sample constitute a second training sample, and then obtaining a plurality of second training samples;

[0123] Step A2: adopt a random sampling strategy with replacement to perform bagging sampling on multiple second training samples to obtain multiple second training samples to form a sub-training set, and then obtain multiple sub-training sets;

[0124] Step A3: for each sub-training set, construct an initial model of a random weight neural network algorithm, wherein the initial model weights and biases are random;

[0125] Step A4: The initial model is trained through a sub-training set. Specifically, the input samples in the sub-training set are input into the initial model to obtain historical prediction values ​​corresponding to historical Gaussian features, wherein the historical prediction values ​​include predicted values ​​of carbon monoxide concentration and predicted values ​​of carbon dioxide concentration. According to the historical prediction values ​​and the output samples, a regularized least squares solution algorithm is used to update the parameters in the initial model until the model converges to obtain a second prediction model, and then multiple different second prediction models corresponding to multiple sub-training sets are obtained.

[0126] In step 5.3, the first weight and the second weight are determined by using the particle swarm optimization algorithm PSO, including:

[0127] Step B1: obtaining a predicted value of carbon monoxide concentration and a true value of carbon monoxide concentration, as well as a predicted value of carbon dioxide concentration and a true value of carbon dioxide concentration, wherein the predicted value of carbon monoxide concentration includes a first concentration value and a second concentration value, and the predicted value of carbon dioxide includes a third concentration value and a fourth concentration value;

[0128] Step B2: Initialize the speed and position of all particles in the particle swarm. The position of each particle includes two randomly initialized weights. The individual optimal position and the global optimal position are initialized, and the initial number of iterations is set.

[0129] Step B3: the initial speed of the particle is used as the current speed, the initial position of the particle is used as the current position, the initial individual optimal position of each particle is used as the current individual optimal position of the particle, the initial global optimal position is used as the current global optimal position, and the initial number of iterations is used as the current number of iterations;

[0130] Step B4: for each particle, according to the weight in the current position of the particle, weighted sum of the first concentration value and the second concentration value is performed to obtain a weighted predicted value of the carbon monoxide concentration, and weighted sum of the third concentration value and the fourth concentration value is performed to obtain a weighted predicted value of the carbon dioxide concentration;

[0131] Step B5: for each particle, the mean square error between the weighted predicted value of carbon monoxide concentration and the true value of carbon monoxide concentration is calculated to obtain a first error, and the mean square error between the weighted predicted value of carbon dioxide concentration and the true value of carbon dioxide concentration is calculated to obtain a second error. The first error and the second error constitute the current fitness of the particle. Among the current fitness of the particle and the fitness corresponding to the current individual optimal position of the particle, the position with the smaller fitness is selected as the new current individual optimal position of the particle.

[0132] Step B6: Among the fitness of the current individual optimal position of all particles after update and the fitness of the current global optimal position, select the position with the smaller fitness as the new current global optimal position;

[0133] Step B7: Update the current position and current velocity of each particle, which is specifically achieved through the following formula:

[0134] v i,t+1 =ωv i,t +c1r1(p best,i -x i,t )+c2r2(g best -x i,t );

[0135] x i,t+1 =x i,t +v i,t+1 ;

[0136] Among them, ω is the inertia weight, c1, c2 are acceleration constants, r1, r2 are random numbers, v i,t+1 is the current velocity of the ith particle at iteration t+1, v i,t is the current velocity of the ith particle at the current iteration number t, p best,i is the current individual optimal position of the i-th particle at the current iteration number t, g best is the current global optimal position of the ith particle at the current iteration number t, x i,t is the current position of the ith particle at the current iteration number t, x i,t+1 is the current position of the i-th particle at iteration t+1;

[0137] Step B8: The current number of iterations is increased by one. When the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold, the two weights in the new current global optimal position are used as the first weight and the second weight. When the current number of iterations is not less than the preset number, and the fitness of the new current global optimal position is not less than the preset threshold, return to execute step B4 until the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold.

[0138] Based on the above technical solution, the present invention has been experimentally verified. The present invention takes the No. 7 blast furnace of a domestic steel plant as the research object and uses actual production record data for research. The proposed RF-RVFLNs combined model is verified by on-site video images, actual production operation parameters of the blast furnace, and the ratio of CO2 and CO in the blast furnace exhaust gas. The 1608 sample data sets after preprocessing are divided into training sets and test sets, of which 1408 are used for training and 200 are used for testing. The ratios of CO2 and CO in the blast furnace exhaust gas used to test the model are as follows: Fig.10 and Fig.11The present invention also attempts other methods, such as a mechanism fusion method based on deep learning to predict blast furnace carbon emission concentration, but these methods are not as accurate as the method proposed in the present invention.

[0139] In order to more intuitively demonstrate that the proposed combined model has higher accuracy and effectiveness, this study uses the test set data to compare the prediction results of the single model and the combined model. The selected comparison algorithms include multi-kernel support vector regression (M-SVR), sparse autoencoder (S-DAE), random forest (RF) and random vector function link network (RVFLNs). These algorithms are used to predict data without furnace top images. The CO2 prediction error is shown in Figure 2. Fig.12 As shown in Figure 2, the CO prediction error is Fig.13 shown.

[0140] Root mean square error (RMSE) and mean absolute percentage error (MAPE) have been used as evaluation metrics and are calculated using the following formulas:

[0141]

[0142] Among them, Table 2 lists the RMSE and MAPE comparison results of different algorithms in CO2 emission concentration prediction, and Table 3 lists the RMSE and MAPE comparison results of CO emission concentration prediction

[0143] Table 2 Comparison of RMSE and MAPE of different algorithms in CO2 emission concentration prediction

[0144] algorithm RMSE(%) MAPE(%) M-SVR 0.6839 2.7921 S-DAE 0.8002 3.1857 RF 0.2956 1.0332 RVFLNS 0.3374 1.7590 RF-RVFLNS 0.2174 0.8742

[0145] Table 3 Comparison results of RMSE and MAPE in CO emission concentration prediction

[0146] algorithm RMSE(%) MAPE(%) M-SVR 0.7626 2.1964 S-DAE 0.6452 1.7748 RF 0.3383 0.9311 RVFLNS 0.3716 1.4461 RF-RVFLNS 0.2844 0.8396

[0147] In summary, the gray-level co-occurrence matrix (GLCM) is used to select the unobstructed flame part from the top thermal image, and ResNet50 and autoencoder are used to extract features from these images. Independent component analysis (ICA) is then applied to extract non-Gaussian independent source features from the image data as the input of the random forest (RF) model. Secondly, the temperature field in the blast furnace is constructed using the mechanism model. The important features of the image are reconstructed using the inverse ICA transform, and they are combined with the operating parameters and temperature field features to establish a feature sharing matrix. Finally, principal component analysis (PCA) is used to extract the principal components as the input of the random vector functional link network (RVFLNs) integration. Finally, particle swarm optimization (PSO) is used to fuse the output results. The results show that the proposed model has a mean absolute percentage error (MAPE) of 0.8742% in terms of CO2 concentration and a mean absolute percentage error of 0.8396% in terms of CO2 concentration, which shows that it has high accuracy and practical application value.

[0148] This paper proposes a carbon emission concentration prediction model based on digital twin technology, which combines video images, operating parameters and temperature field features. By adopting independent component analysis (ICA) and principal component analysis (PCA) algorithms for feature fusion, and using the random forest (RF)-extreme learning machine (RVFLNs) integrated model for prediction, the experimental results show that this method significantly improves the accuracy and stability of carbon emission prediction in an actual production environment. It effectively utilizes the advantages of each model and overcomes the limitations of a single model. However, there are still some limitations in this study. During the image feature extraction process, some local features may be lost. In the future, it can be studied how to directly use the image matrix for regression learning while ensuring real-time performance.

[0149] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that this should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to), and the scope of the invention involved in the disclosed embodiments is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features.

Claims

1. A method for predicting carbon emission concentration of blast furnace based on digital twin, characterized in that: include: Step 1: Real-time collection of production data during blast furnace smelting, wherein the production data includes operating parameters, tuyere parameters and multiple blast furnace top thermal images, wherein the tuyere parameters include tuyere temperature and tuyere velocity; Step 2: Screen all the thermal images of the blast furnace top to obtain the thermal images that are not blocked by the charging chute; Step 3: According to the thermal imaging image and tuyere parameters not blocked by the charging chute, the temperature field characteristics are obtained, wherein the temperature field characteristics include temperature gradient, maximum temperature and average temperature; Step 4: For each thermal image that is not blocked by the charging chute, extract the features through ResNet50 and autoencoder to obtain the image features of all thermal images that are not blocked by the charging chute, and compress them into an image feature matrix X. The independent component analysis algorithm is used to extract the image feature matrix X to obtain the non-Gaussian independent source feature S. The non-Gaussian independent source feature S is processed by ICA inverse transformation to obtain the image key feature X. e , the key feature X of the image e Combined with the physical variable Z, we get the shared matrix X c =[Z;X e ], where the physical variable Z includes the operating parameters and temperature field characteristics after preprocessing, and the principal component analysis algorithm is used to extract the Gaussian features in the shared matrix; Step 5: Obtain the final carbon monoxide concentration and carbon dioxide concentration based on the non-Gaussian independent source feature S and the Gaussian feature.

2. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 1 is characterized in that: Step 2 specifically includes: Through the gray level co-occurrence matrix, feature extraction is performed on each of the blast furnace top thermal images to obtain the texture features of each blast furnace top thermal image, wherein the texture features include contrast, energy, uniformity and the difference of pixels in the blast furnace top thermal image. According to the blast furnace top thermal image and the corresponding texture features, all the blast furnace top thermal images are screened to obtain the thermal images that are not blocked by the charging chute.

3. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 2 is characterized in that: According to the blast furnace top thermal image and the corresponding texture features, all blast furnace top thermal images are screened to obtain thermal images that are not blocked by the charging chute, including: According to the thermal imaging image of the blast furnace top, the production stage of the blast furnace is determined. When the blast furnace is in the unloading stage, the value range of texture features not blocked by the charging chute corresponding to the unloading stage is obtained, and it is used as the target value range. When the blast furnace is in the charging stage, the value range of texture features not blocked by the charging chute corresponding to the charging stage is obtained, and it is used as the target value range; obtain the thermal imaging image of the blast furnace top with texture features within the target value range, and obtain the thermal imaging image not blocked by the charging chute.

4. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 1 is characterized in that: Step 3 specifically includes: The mass conservation equation, momentum conservation equation and energy conservation equation are used to process the tuyere parameters to obtain the initial temperature field. For the flame area in the thermal imaging image that is not blocked by the charging chute, the grayscale value of the flame area is extracted, and the actual temperature value corresponding to the grayscale value is obtained. The actual temperature value is combined with the initial temperature field to obtain the final temperature field. Feature extraction is performed on the final temperature field to obtain the temperature field characteristics.

5. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 1 is characterized in that: Step 5 specifically includes: Step 5.1: Process the non-Gaussian independent source feature S in step 4 by using a first prediction model based on a random forest regression algorithm to obtain a first concentration prediction value of carbon monoxide and a first concentration prediction value of carbon dioxide; The first prediction model is obtained by model training based on multiple first training samples, the first training samples include a first input sample and a first output sample, the first input sample is a historical non-Gaussian independent source feature, and the first output sample includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration corresponding to the historical non-Gaussian independent source feature; Step 5.2: The Gaussian features in step 4 are processed by a plurality of different second prediction models based on a random weight neural network algorithm to obtain a plurality of second prediction values, wherein the second prediction values ​​include a second concentration prediction value of carbon monoxide and a second concentration prediction value of carbon dioxide. According to a preset weight, all the second concentration prediction values ​​of carbon monoxide are weighted averaged to obtain a third concentration prediction value of carbon monoxide, and all the second concentration prediction values ​​of carbon dioxide are weighted averaged to obtain a third concentration prediction value of carbon dioxide. Step 5.3: Determine the first weight and the second weight through the particle swarm optimization algorithm PSO, and perform weighted summation on the first predicted value of carbon monoxide concentration and the third predicted value of carbon monoxide concentration according to the first weight and the second weight to obtain the final carbon monoxide concentration, and perform weighted summation on the first predicted value of carbon dioxide concentration and the third predicted value of carbon dioxide concentration according to the first weight and the second weight to obtain the final carbon dioxide concentration.

6. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 5 is characterized in that: In step 5.3, a plurality of different second prediction models are obtained by the following steps: Step A1: obtaining a historical Gaussian feature and a true concentration value corresponding to the historical Gaussian feature, wherein the true concentration value includes a true value of carbon monoxide concentration and a true value of carbon dioxide concentration, taking the historical Gaussian feature as a second input sample, taking the true concentration value as a second output sample, the second input sample and the second output sample constitute a second training sample, and then obtaining a plurality of second training samples; Step A2: adopt a random sampling strategy with replacement to perform bagging sampling on multiple second training samples to obtain multiple second training samples to form a sub-training set, and then obtain multiple sub-training sets; Step A3: for each sub-training set, construct an initial model of a random weight neural network algorithm, wherein the initial model weights and biases are random; Step A4: The initial model is trained through a sub-training set. Specifically, the input samples in the sub-training set are input into the initial model to obtain historical prediction values ​​corresponding to historical Gaussian features, wherein the historical prediction values ​​include predicted values ​​of carbon monoxide concentration and predicted values ​​of carbon dioxide concentration. According to the historical prediction values ​​and the output samples, a regularized least squares solution algorithm is used to update the parameters in the initial model until the model converges to obtain a second prediction model, and then multiple different second prediction models corresponding to multiple sub-training sets are obtained.

7. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 5 is characterized in that: In step 5.3, the first weight and the second weight are determined by using the particle swarm optimization algorithm PSO, including: Step B1: obtaining a predicted value of carbon monoxide concentration and a true value of carbon monoxide concentration, as well as a predicted value of carbon dioxide concentration and a true value of carbon dioxide concentration, wherein the predicted value of carbon monoxide concentration includes a first concentration value and a second concentration value, and the predicted value of carbon dioxide includes a third concentration value and a fourth concentration value; Step B2: Initialize the speed and position of all particles in the particle swarm. The position of each particle includes two randomly initialized weights. The individual optimal position and the global optimal position are initialized, and the initial number of iterations is set. Step B3: the initial speed of the particle is used as the current speed, the initial position of the particle is used as the current position, the initial individual optimal position of each particle is used as the current individual optimal position of the particle, the initial global optimal position is used as the current global optimal position, and the initial number of iterations is used as the current number of iterations; Step B4: for each particle, according to the weight in the current position of the particle, weighted sum of the first concentration value and the second concentration value is performed to obtain a weighted predicted value of the carbon monoxide concentration, and weighted sum of the third concentration value and the fourth concentration value is performed to obtain a weighted predicted value of the carbon dioxide concentration; Step B5: for each particle, the mean square error between the weighted predicted value of carbon monoxide concentration and the true value of carbon monoxide concentration is calculated to obtain a first error, and the mean square error between the weighted predicted value of carbon dioxide concentration and the true value of carbon dioxide concentration is calculated to obtain a second error. The first error and the second error constitute the current fitness of the particle. Among the current fitness of the particle and the fitness corresponding to the current individual optimal position of the particle, the position with the smaller fitness is selected as the new current individual optimal position of the particle. Step B6: Among the fitness of the current individual optimal position of all particles after update and the fitness of the current global optimal position, select the position with the smaller fitness as the new current global optimal position; Step B7: Update the current position and current speed of each particle, which is specifically achieved through the following formula: v i,t+1 =ωv i,t +c1r1(p best,i -x i,t )+c2r2(g best -x i,t x i,t+1 =x i,t +v i,t+1 Among them, ω is the inertia weight, c1, c2 are acceleration constants, r1, r2 are random numbers, v i,t+1 is the current velocity of the ith particle at iteration t+1, v i,t is the current velocity of the ith particle at the current iteration number t, p best,i is the current individual optimal position of the i-th particle at the current iteration number t, g best is the current global optimal position of the ith particle at the current iteration number t, x i,t is the current position of the ith particle at the current iteration number t, x i,t+1 is the current position of the i-th particle at iteration t+1; Step B8: The current number of iterations is increased by one. When the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold, the two weights in the new current global optimal position are used as the first weight and the second weight. When the current number of iterations is not less than the preset number, and the fitness of the new current global optimal position is not less than the preset threshold, return to execute step B4 until the current number of iterations is less than the preset number, or the fitness of the new current global optimal position is less than the preset threshold.

8. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 1 is characterized in that: The independent component analysis algorithm is used to extract the image feature matrix X to obtain the non-Gaussian independent source feature S, which is specifically implemented by the following formula: S = WX; Where W is the unmixing matrix.

9. The method for predicting carbon emission concentration of a blast furnace based on digital twin according to claim 1 is characterized in that: The ICA inverse transform is used to process the non-Gaussian independent source feature S to obtain the key feature X of the image. e , which is specifically achieved through the following formula: X e＝ X-AS=(I-AW)X; Among them, A is the mixing coefficient matrix and I is the unit matrix.

Images