Blast furnace carbon emission anomaly prediction method based on digital twinning

Abstract

The present invention provides a method for predicting abnormal carbon emissions from blast furnaces based on digital twins, which relates to the technical field of blast furnace carbon emissions. The present invention proposes a method for detecting abnormal carbon emissions from the top of a blast furnace based on deep learning and mechanism fusion. First, the key features of carbon traces in the blast furnace production process are collected, and the gas composition and peripheral production data at the top of the blast furnace are collected through sensors. The characteristic data that some sensors cannot obtain in real time are iteratively solved by the Navier-Stokes NS equation. Then, cross-correlation analysis is performed to discover the lag relationship of carbon traces in production, and a carbon emission lag relationship model is established. Abnormal operating conditions are detected by a residual neural network. Finally, the performance of the model is tested and evaluated using actual production data. The results show that the proposed method can accurately detect abnormal carbon emissions from the top of the blast furnace, and can issue early warnings in a timely manner, effectively reducing environmental pollution and energy waste caused by abnormal carbon emissions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blast furnace carbon emission, and particularly relates to a blast furnace carbon emission anomaly prediction method based on digital twinning. BACKGROUND

[0002] Carbon emission in the process of blast furnace ironmaking is an important source of greenhouse gas emission in steel production. Effective detection and control of abnormal carbon emission at the top of the blast furnace is of great significance for reducing carbon emission, improving energy efficiency and achieving green steel production. The current blast furnace gas data-driven method has high data dependency and low interpretability.

[0003] As the core equipment of the steel production process, the blast furnace is responsible for the key task of converting iron ore into molten iron. The blast furnace reaction heap is large in scale, and the smelting process is carried out under high temperature and high pressure. This is a core challenge in steel smelting technology. The stable operation of the blast furnace is not only crucial to the continuity and efficiency of steel production, but also an important symbol of the technical level and competitiveness of steel enterprises. In this case, the importance of blast furnace anomaly prediction becomes particularly evident. As a key technical means, anomaly prediction can identify and warn potential abnormal hazards in time through real-time monitoring and comprehensive analysis of the operating conditions of the blast furnace, thereby providing a scientific basis for production management and maintenance decision-making. This not only helps to avoid production interruptions and safety accidents, reduces economic losses and personnel casualties, but also optimizes blast furnace operating parameters, improves production efficiency and product quality, and promotes the technological progress and industrial upgrading of the steel industry. Therefore, the in-depth study and application of blast furnace anomaly prediction technology has great production value and practical significance for ensuring the stability, safety and efficiency of steel production.

[0004] In the process of blast furnace ironmaking, various types of faults may occur. According to the location and nature of the fault, it can be divided into abnormal furnace condition fault, equipment fault and raw material quality fault. Among them, the abnormal furnace condition fault is the most threatening to production. This abnormality mainly manifests as uneven gas flow distribution in the blast furnace, large temperature fluctuations and poor descent of the furnace charge. Specifically, it includes suspended material, material collapse, pipe accumulation and hearth accumulation, etc. Due to the complexity of multiphase fluid mechanics in the blast furnace and the uncertainty of internal chemical reactions, it is difficult to accurately and quantitatively analyze the physical variables such as velocity and temperature of the components and materials at each position in the blast furnace. Therefore, it is difficult to quickly establish an accurate mechanism model to describe the production process of blast furnace ironmaking. In the production of blast furnace, when abnormal conditions occur, it will cause serious fluctuations in the operating conditions of the blast furnace, which will seriously affect the blast furnace output, molten iron quality and production energy consumption.

[0005] After years of unremitting efforts by researchers, many methods for blast furnace anomaly detection have been developed. Early research methods rely more on expert experience and sensors, such as the analytic hierarchy process (AHP), technique for order preference by similarity to ideal solution (TOPSIS), and Bayesian networks, to establish evaluation models and obtain better evaluation results. However, relying on expert experience can be costly and has certain limitations. With the improvement of computing power, data-driven models have become an important method for blast furnace anomaly detection. For example, support vector machines (SVM), kernel correlation filter (KCF), and other machine learning methods. Zhou Ping et al. proposed a novel integrated PCA-ICA method to monitor and diagnose blast furnace anomalies. Despite extensive research, there are still some unresolved issues in the establishment of blast furnace ironmaking process monitoring models. Most of the above models have relatively simple structures and perform well when dealing with simple data structures. When the data structure is complex, shallow models perform poorly in capturing complex nonlinear relationships between data. Wu Ping et al. used a deep learning model for blast furnace anomaly prediction and proposed an anomaly prediction method based on multi-channel dynamic GCN (MDGCN). In the graph, different nodes are assigned weights to extract more useful information about the process dynamics for fault diagnosis. This method has attracted widespread attention and has been applied in many fields. Feature extraction plays a crucial role in deep learning models. Raw data often contains a lot of redundant information. Directly processing these high-dimensional data not only requires a lot of computation, but also may cause model overfitting. Feature extraction extracts key information from data and converts it into a low-dimensional but more representative feature vector, thereby reducing computational cost and improving model generalization ability. An efficient feature extraction method can improve the accuracy of anomaly detection. Lu Siwei et al. proposed a local-dynamic generalized kernel stationary subspace analysis method (local-dbkssa) that has higher robustness. Feature extraction can also use techniques such as encoders, convolutional neural networks (CNN), and recurrent neural networks. Huang Hang et al. proposed an anomaly prediction and identification method based on a multi-dimensional gated recurrent unit (GRU) network, which has a simpler structure and fewer parameters but still provides good prediction results. Although these methods show high precision in blast furnace state anomaly detection, relying solely on data-driven models for fault identification has drawbacks such as poor generalization ability and long training time, which cannot meet the speed requirements of blast furnace production. From the perspective of blast furnace carbon emissions, the main source of carbon dioxide gas emissions is coke and sintered ore. Coke and sintered ore enter the blast furnace through the distribution clock at the top, undergo complex chemical reactions, and are converted into blast furnace gas and subsequently discharged. This reaction process takes a long time. When using data-driven models for anomaly prediction, the significant lag in the production process and the impact of internal reactions on anomaly prediction are often ignored. Relying solely on the nonlinear mapping ability of deep learning may slightly reduce accuracy.Currently, fluid mechanics is developing rapidly. By using computational fluid dynamics technology, the velocity field and temperature field in the blast furnace can be quantitatively analyzed to extract additional features, which helps to analyze the abnormality of the blast furnace.

[0006] Digital twin technology is a key advancement in digital transformation and intelligent upgrading. Driven by data and models, digital twin technology simulates the dynamic response of physical entities to internal and external mechanisms, and can perform monitoring, simulation, correlation, prediction, optimization, and other operations. As a core component of steelmaking, the chemical reactions inside the blast furnace are both complex and critical. Among these reactions, carbon dissolution reaction, direct reduction reaction, indirect reduction reaction, and internal reduction reaction are crucial. These reactions not only affect the quality and yield of the blast furnace molten iron, but also reveal the reaction state inside the blast furnace through the composition of the blast furnace gas emitted at the top of the blast furnace.

[0007] The carbon dioxide content in the blast furnace gas is a key indicator of the reactor's condition, which is influenced by various factors, including raw material quality, operational parameter adjustments, and changes in equipment status. Any abnormal fluctuations in carbon monoxide content can adversely affect the operational efficiency of the blast furnace and potentially jeopardize the normal operation of the blast furnace and subsequent process equipment, thereby impacting the economic and operational stability of the entire steel production line. Given the importance of this issue, it is crucial to promptly and accurately diagnose abnormal carbon monoxide content in the blast furnace gas and take appropriate measures to adjust and control it. This process not only ensures the smooth operation of the blast furnace, but also significantly improves production efficiency, reduces energy consumption, and reduces harmful gas emissions.

[0008] In the blast furnace production process, the carbon sources mainly include ore and coke added at the top, and coal powder injected through the tuyere at the bottom. These carbon sources undergo high temperature and high pressure in the furnace and undergo a series of reduction reactions with iron ore, ultimately producing molten iron and coal gas. However, there is a significant delay between the top of the furnace and the completion of the reaction and the emission of blast furnace gas. This delay poses a major challenge to analyzing the reaction state inside the furnace. By detecting and analyzing the proportions of various gas components in the blast furnace gas, we can more accurately assess whether the reduction reactions inside the blast furnace are sufficient and monitor changes in key parameters such as temperature and pressure. SUMMARY

[0009] To address the shortcomings of the prior art, the present application provides a blast furnace carbon emission anomaly prediction method based on digital twinning. The blast furnace operating parameters, such as ore addition amount and coke addition rate, are used to predict the carbon emissions at the top of the blast furnace. By precisely controlling these parameters and combining the professional knowledge of on-site personnel, we can achieve real-time monitoring and dynamic adjustment of the blast furnace operating state. Once any abnormal reaction or signs of failure are detected in the blast furnace, we can quickly take appropriate intervention measures to prevent failure and ensure stable operation and efficient production of the blast furnace.

[0010] A blast furnace carbon emission anomaly prediction method based on digital twinning, comprising the following steps:

[0011] Step 1: Construct a digital twinning mechanism model to simulate and monitor the changes of velocity field, temperature field and pressure field in real time; Specifically, solve the Navier-Stokes NS equation of two-dimensional incompressible fluid by using digital twinning technology; Determine the velocity field, temperature field and pressure field by solving the Navier-Stokes NS equation and the heat conduction differential equation;

[0012] Step 1.1: Generate a blast furnace modeling area, specifically use an elliptic partial differential equation to generate a grid for the modeling area;

[0013] The arrangement of the grid is consistent with the longitudinal direction of the blast furnace;

[0014] Step 1.2: Assume that the fluid in the blast furnace is incompressible, simplify the NS equation to two-dimensional form; Obtain the velocity field and pressure field inside the blast furnace by iteratively solving the discretized two-dimensional NS equation; At the same time, use the temperature value obtained by the infrared thermocouple in the blast furnace tuyere rotation zone to solve the heat differential equation to determine the temperature field; The temperature field, pressure field and velocity field constitute the mechanism model.

[0015] Step 1.3: Establish a digital twinning data-driven model;

[0016] Step 1.3.1: Take edge sampling points from the mechanism model and fuse with real-time sensors to form input data for the data-driven model;

[0017] Step 1.3.2: Use cross-correlation analysis to determine the carbon time lag value under the current working condition;

[0018] The cross-correlation analysis is as follows:

[0019] Through cross-correlation analysis, the lag relationship between the blast furnace top distribution data and the blast furnace top gas composition data is determined, and the formula is as follows:

[0020]

[0021] In the formula, x(i) is the input amount of carbon, y(i) is the output of carbon, N is the time span of the data, R xy (k) is the result of cross-correlation analysis;

[0022] Use fast Fourier transform to convert carbon correlation data from time domain to frequency domain, and then convert back to time domain;

[0023] R xy(k) = ifft(fft(x).* conj(fft(y))) (2)

[0024] In the formula, fft is fast Fourier transform, ifft is inverse fast Fourier transform, conj is conjugate;

[0025] Through cross-correlation analysis, the carbon lag time of the blast furnace under the current working condition is obtained; the top pressure, temperature and edge detection data are taken as the input data of cross-correlation, and the historical data of the top gas composition are taken as the target data, that is, x and y in the cross-correlation analysis; based on the lag time under the current working condition obtained through cross-correlation analysis, the input data and the output data are matched, the PCA dimension reduction method is adopted, and the order after dimension reduction is taken as the input of the data-driven model;

[0026] Step 1.3.3: Constructing a data-driven model, that is, a residual neural network;

[0027] After obtaining the carbon lag time under the current working condition, deep learning is used to predict the carbon emission anomaly; specifically, an optimized convolutional neural network CNN is used to diagnose the carbon emission anomaly under various working conditions; the optimized convolutional neural network CNN takes feature input as the starting point, first passes through two groups of parallel convolution operations, each group containing 32 convolution kernels, and then connects to the ReLU activation function, the feature map is fused through addition and then enters the next stage, then the feature map is reduced in size through the pooling layer and preprocessed through the 0 padding method; in the next stage, the number of convolution layers is expanded to 64, and the activation function is continued to be used for processing, the features are fused and then enter the final fully connected layer to complete feature extraction and mapping; the whole network realizes layer-by-layer refining and information aggregation of features through multi-level convolution, pooling and full connection operations, and the calculation method of output vector y is as follows:

[0028] y = H(x) = F(x, {W i}) + x (3)

[0029] Where F(x, {Wi}) represents the learned residual mapping, and x is the input vector;

[0030] Step 1.4: Data preprocessing is performed on the data-driven model to obtain a digital twin mechanism model; specifically, the data is normalized; the normalization formula is applied to scale the data to the range of [-1, 1]:

[0031] x norm =2*(x-x min ) / (x max -x min )-1 (4)

[0032] In the formula, x represents the input data that needs to be preprocessed, x norm represents the normalized result, and x min and x maxmin and max, respectively, denote the minimum and maximum of a set of variables;

[0033] Step 2: After obtaining the sensor data that can be obtained in real time at the current moment, construct the velocity field, pressure field, and temperature field model, and fuse with the sensor data as the input data of the data-driven model, and obtain the time lag in the short-term working condition through cross-correlation analysis. After inputting the data into the data-driven model, the amount of carbon emission at the top of the furnace is obtained, combined with the time lag, to determine the time of the predicted carbon emission amount.

[0034] The beneficial effects produced by the above technical solutions are:

[0035] The present application provides a blast furnace carbon emission anomaly prediction method based on digital twinning. In blast furnace ironmaking, the detection of abnormal state is crucial for reducing accident risk, improving economic benefit, and realizing safe and sustainable production. The present application proposes a digital twinning mechanism model, i.e. variables that cannot be obtained in real time, and then combines these variables with real-time sensor data. Cross-correlation analysis is used to determine the time lag of carbon emission under various operating conditions. Then, a deep neural network is constructed using residual mapping for detecting abnormal carbon emission at the top of the furnace. The results of the verification using production data from a steel plant in this embodiment show that this method can accurately detect the insufficient reduction conditions in the ironmaking process. In order to improve the detection accuracy, we will further improve the mechanism part, establish the internal composition field of the blast furnace, and improve the accuracy of anomaly prediction. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a thermal imaging image of the top of the blast furnace of the present application;

[0037] Figure 2 is a whole flow chart of the blast furnace carbon emission anomaly prediction method of the present application;

[0038] Figure 3 is a flow chart of the modeling area generation method of the present application;

[0039] Figure 4 is an infrared temperature measurement result graph of the tuyere rotation zone of the present application;

[0040] Figure 5 is a lag relationship curve of carbon emission of the present application;

[0041] Figure 6 is a residual neural network architecture diagram of the present application;

[0042] Figure 7 is a skip connection schematic diagram of the present application;

[0043] Figure 8 is a comparison schematic diagram of algorithm performance in prediction of the present application;

[0044] Figure 9 Scatter plot for the regression relationship of the present application;

[0045] Figure 10 Histogram for the error distribution in the display data of the present application;

[0046] Figure 11 Confusion matrix plot for anomaly detection of the present application;

[0047] Figure 12 Cumulative error curve plot of the present application. DETAILED DESCRIPTION

[0048] The specific embodiments of the present application will be further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application, but are not used to limit the scope of the present application.

[0049] A blast furnace carbon emission anomaly prediction method based on digital twinning, as shown in Figure 2 , includes the following steps:

[0050] Step 1: Construct a digital twinning mechanism model to simulate and monitor the changes of velocity field, temperature field and pressure field in real time; Specifically, solve the Navier-Stokes NS equation of two-dimensional incompressible fluid by using digital twinning technology; Digital twinning technology not only improves the accuracy of simulation, but also helps us better understand and predict the running status of the system. By solving the Navier-Stokes NS equation and the heat conduction differential equation, the velocity field, temperature field and pressure field are determined;

[0051] Step 1.1: Generate a blast furnace modeling area, specifically use elliptic partial differential equations to generate the grid of the modeling area;

[0052] In this embodiment, the blast furnace is 27.1 meters high, and the radius of the maximum cross section is 6.6 meters. The thermal imaging diagram of the top of the blast furnace is as shown in Figure 1 , Figure 3 shows the iteration process of the grid of the modeling area, and this embodiment selects to divide the modeling area radially into 60 steps and vertically into 271 steps, a total of 16,260 grids are created. The average step length of all grids is about 100 millimeters, which can accommodate multiple particles, reduce the delay of calculation, and at the same time ensure the accuracy of the model. The arrangement of the grid is consistent with the longitudinal direction of the blast furnace; according to experience, the gas in the blast furnace flows from the bottom to the top, and its flow trajectory is consistent with the grid arrangement, which helps the model to converge.

[0053] The NS equations are the fundamental mathematical tools for describing fluid motion. These equations describe the motion of fluids and are crucial in various engineering and scientific fields. They reveal the essence of fluid motion: the fluid velocity changes under the influence of external forces (such as pressure gradients, body forces) and internal forces (such as viscous stresses). This change is not only related to the current velocity and pressure distribution but also to the physical properties of the fluid (such as density, viscosity) and the flow history (such as convection terms). Therefore, the NS equations provide a powerful tool for studying the laws of fluid flow. The three conservation laws contained in the NS equations are shown in Figure 3 .

[0054] Step 1.2: Currently, there is no analytical solution for the three-dimensional compressible fluid NS equations. The process of using discrete methods for iterative calculations to find approximate solutions is extremely slow and cannot meet the requirements of industrial calculation speed. Assuming that the fluid in the blast furnace is incompressible, the NS equations are simplified to two-dimensional form; by iteratively solving the discretized two-dimensional NS equations, the velocity field and pressure field inside the blast furnace are obtained; at the same time, the temperature field is determined by solving the heat differential equation using the temperature values obtained from the infrared thermocouple in the blast furnace tuyere swirl zone; at this stage, the temperature field, pressure field, and velocity field, which cannot be measured in real time, constitute the mechanism model.

[0055] Step 1.3: Establish a digital twin data-driven model;

[0056] Step 1.3.1: Take edge sampling points from the mechanism model and fuse with real-time sensors to form input data for the data-driven model; including furnace top pressure sensors, tuyere swirl zone infrared temperature measuring instruments, etc.; the infrared temperature measurement results in the tuyere swirl zone are shown in Figure 4 .

[0057] Step 1.3.2: To solve the carbon lag problem, cross-correlation analysis is used to determine the carbon lag time lag value under the current working condition;

[0058] The cross-correlation analysis is as follows:

[0059] Due to the long time lag between the addition and discharge of carbon in blast furnace production, there is a time difference when sampling and analyzing the gas in the blast furnace. As shown in Figure 5 , it is crucial to determine the time required for carbon to go from addition to discharge. From a steel plant, obtain the blast furnace top distribution data and blast furnace top gas composition data. Through cross-correlation analysis, determine the lag relationship between the blast furnace top distribution data and the blast furnace top gas composition data, as follows:

[0060]

[0061] In the formula, x(i) is the input amount of carbon, y(i) is the output of carbon, N is the time span of the data, R xy(k) is the result of the cross-correlation analysis.

[0062] Since all the time difference points need to be nested in two layers of for loops, the calculation load is particularly large in scale. Therefore, the carbon correlation data is converted from time domain to frequency domain using fast Fourier transform, and then converted back to time domain to simplify the operation.

[0063] R xy(k) = ifft(fft(x).*conj(fft(y))) (2)

[0064] In the formula, fft is the fast Fourier transform, ifft is the inverse fast Fourier transform, and conj is the conjugate;

[0065] Through cross-correlation analysis, the carbon lag time of the blast furnace under the current working condition is obtained; the top pressure, temperature and edge detection data are taken as the input data of cross-correlation, and the historical data of the top gas composition are taken as the target data, that is, x and y in the cross-correlation analysis; based on the lag time under the current working condition obtained through the cross-correlation analysis, the input data and the output data are matched, in order to meet the requirement of industry on speed, reduce the training components and make the model converge faster, the PCA dimension reduction method is adopted, and the order after dimension reduction is taken as the input of the data-driven model;

[0066] Step 1.3.3: Construct a data-driven model, that is, a residual neural network, as shown in Figure 6 、 Figure 7 ;

[0067] After obtaining the carbon lag time under the current working condition, deep learning is used to predict carbon emission anomalies; inspired by residual neural networks, in the residual block, the input can be transmitted faster through the cross-layer data line. In addition, through the jump connection, the signal can be transmitted without attenuation during the back propagation process, alleviating the problem of gradient disappearance or gradient explosion caused by deepening the number of layers. It can also reduce the occurrence of overfitting of deep networks. Specifically, an optimized convolutional neural network CNN is used to diagnose carbon emission anomalies under various working conditions; the optimized convolutional neural network CNN starts with feature input, first passes through two groups of parallel convolution operations, each group containing 32 convolution kernels, and then connects to the ReLU activation function. After the feature map is fused by addition, it enters the next stage, and then the feature map is reduced in size through the pooling layer and adjusted in size through the 0 padding method to complete the preprocessing; in the next stage, the number of convolution layers is expanded to 64, and the activation function is still used for processing. After the features are fused, they enter the final fully connected layer to complete feature extraction and mapping; the entire network realizes layer-by-layer refinement and information aggregation of features through multi-level convolution, pooling and full connection operations, and has efficient feature learning ability. The training accuracy of the model with too many convolution layers is lower than that of the three-layer convolution model; increasing the number of layers will increase the training error. Compared with using multiple network layers to fit the hidden nonlinear mapping, the network is easier to learn its residual. This premise constitutes the basic concept of the residual block. The basic structure of the residual block is as follows:

[0068] y=H(x)=F(x,{W i})+x (3)

[0069] Where F(x, {W i}) represents the learned residual mapping, and x is the input vector; in the two-layer structure shown in Figure 7 , the dimensions of x and F must be equal when performing summation. When the number of input or output channels changes, the resize module size step is adjusted to adjust the dimensions of the data to match their sizes, as shown in Figure 7 . In this embodiment, a uniform kernel size of 3x1 is used, and the number of channels and network structure are as shown in Figure 6 .

[0070] Step 1.4: Data-driven model data preprocessing; industrial data is often affected by noise and outliers, and data preprocessing is needed to alleviate these effects. In order to improve the accuracy of the model and speed up the convergence speed, the data must be preprocessed. Specifically, the data is normalized; the normalization formula is applied to scale the data to the range of [-1, 1]:

[0071] x norm =2*(x-x min ) / (xmax x min )-1 (4)

[0072] where x represents the input data that needs to be preprocessed, x norm represents the normalized result, and x min and x max represent the minimum and maximum values of a variable set, respectively. In blast furnace production, different variables have different effects on operating conditions. Using all variables can lead to slow model convergence and increased risk of overfitting. Therefore, in this embodiment, principal component analysis (PCA) is used for feature selection.

[0073] Step 2: In blast furnace production, the input and output of carbon have a lag. After obtaining the real-time sensor data at the current time, the velocity field, pressure field, and temperature field models are constructed and fused with the sensor data as input data for the data-driven model. The lag time within the short-term operating condition is obtained through cross-correlation analysis. After inputting the data into the data-driven model, the amount of carbon emissions at the top of the furnace is obtained. Combined with the lag time, the time of the predicted carbon emissions is determined.

[0074] In this embodiment, more than 1,000 sets of production data were selected from an experimental blast furnace in a steel plant, with a sampling frequency of once per hour. After preprocessing, the previous 1,000 hours of data were selected and divided into a training set and a test set in a 17:3 ratio. The test set included 6 small anomalies, 4 medium anomalies, and 6 large anomalies. To speed up model convergence, the network structure was designed with only two residual mappings. The Xavier initializer generates weights with normal distribution, zero mean, and specific variance. The activation function used is ReLU, and the output is generated through a fully connected layer. The gradient descent algorithm used is Adam, with a maximum of 500 training rounds, a minimum batch size of 300, and a learning rate adjustment every 300 rounds. The learning rate adjustment coefficient is 0.2.

[0075] To demonstrate the effectiveness of the proposed method, we conducted comparative experiments with GRU, CNN, SVM, random forest, and resnet18. The GRU algorithm performs single-step prediction, while the CNN uses two convolutional layers. All methods were implemented in MATLAB 2023a and ran on a laboratory computer equipped with an Intel Core i7-13700KF CPU, an RTX 4060Ti GPU, a maximum frequency of 5.4 GHz, and 32 GB of memory. Precision was measured by averaging the results of five experiments.

[0076] Figure 8The performance of different algorithms is compared in Table 1 and Table 2 to evaluate their effectiveness in the prediction task. The results show that time series prediction algorithms perform poorly on chaotic industrial production data, achieving good results only in predicting the next immediate step.

[0077] However, they still fall short compared to convolutional neural network algorithms. Signal decomposition algorithms, such as Empirical Mode Decomposition (EMD), show good results, but obtaining sufficient data for such decomposition is challenging in real production scenarios. The accuracy of convolutional neural networks decreases as the network structure deepens. Fewer convolutional layers result in insufficient non-linear mapping capabilities, leading to inaccurate fault point regression. Resnet addresses the challenge of training deep neural networks by introducing residual mapping. It exhibits strong non-linear mapping capabilities and performs well under some stable working conditions. However, when production inputs change significantly, short-term faults cannot be accurately predicted. Experiments on existing ResNet-50 and ResNet-18 models show poor performance and long training times. ResNet-50 has ten times longer training time than other models, while ResNet-18 also experiences training time extension and poor performance. In the network structure described in this paper, increasing or decreasing the number of network layers leads to a decrease in accuracy. Figure 9 The scatter plot shown demonstrates the regression relationship.

[0078] Table 1: Precision and computation of different methods

[0079]

[0080] Table 2: Accuracy of anomaly prediction

[0081]

[0082] It demonstrates the correlation between the predicted variable and the actual value. Figure 10 A histogram of error distribution is shown, which helps analyze the frequency and distribution of errors. Figure 11 A confusion matrix for anomaly detection is provided, clearly showing the classification accuracy and performance of the model on different classes. Figure 12 A cumulative error curve is depicted, showing the accumulation of errors and stability as the amount of data increases.

[0083] From the above analysis, the residual mapping neural network based on cross-correlation analysis is obviously superior to other methods. Cross-correlation analysis establishes a link between carbon input and output, enhances the consistency between model input and output, improves the interpretability of the neural network, and reduces the nonlinearity of input and output data compared with direct input of different operating conditions. This makes the network regression more accurate. In addition, the introduction of residual mapping enhances the feature extraction capability of the model, making the accuracy of the model higher compared with other algorithms.

[0084] The above description is only the preferred embodiments of the present disclosure and the explanation of the principles of the technology applied. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or equivalent features without departing from the above inventive concept. For example, the above features are replaced with each other to form a technical solution with similar functions disclosed in the embodiments of the present disclosure (but not limited to).

Claims

1. A method for predicting abnormal carbon emissions from blast furnaces based on digital twins, characterized in that: The following steps are involved: Step 1: Build a digital twin mechanism model to simulate and monitor changes in velocity, temperature, and pressure fields in real time; The step 1 specifically includes the following steps: Step 1.1: Generate the blast furnace modeling area. Specifically, use elliptic partial differential equations to generate the grid of the modeling area. The arrangement of the grids is consistent with the longitudinal direction of the blast furnace; Step 1.2: Assuming the fluid in the blast furnace is incompressible, simplify the NS equations to a two-dimensional form. Obtain the velocity and pressure fields inside the blast furnace by iteratively solving the discretized two-dimensional NS equations. Simultaneously, use the temperature values ​​obtained by infrared thermocouples in the blast furnace tuyere raceway to solve the thermal differential equation to determine the temperature field. The temperature field, pressure field, and velocity field constitute the mechanism model. Step 1.3: Build a digital twin data-driven model; The step 1.3 includes the following steps: Step 1.3.1: Take edge sampling points from the mechanism model and fuse them with real-time sensors to form the input data of the data-driven model; Step 1.3.2: Use cross-correlation analysis to determine the carbon lag time lag value under the current operating conditions; The cross-correlation analysis described in step 1.3.2 is as follows: Through cross-correlation analysis, the lag relationship between the blast furnace top distribution data and the blast furnace top gas composition data is determined, and the formula is as follows: Where x(i) is the carbon input, y(i) is the carbon output, N is the time span of the data, and R xy (k) is the result of cross-correlation analysis; The carbon-related data were converted from the time domain to the frequency domain and then back to the time domain using fast Fourier transform; R xy(k) =ifft(fft(x).*conj(fft(y))) (2) Where, fft is the fast Fourier transform, ifft is the inverse fast Fourier transform, and conj is the conjugate. Through cross-correlation analysis, the carbon lag time of the blast furnace under the current operating conditions is obtained. The furnace top pressure, temperature, and edge detection data are used as the input data for the cross-correlation, while the historical data of the furnace top gas composition is used as the target data, i.e., x and y in the cross-correlation analysis. Based on the lag time under the current operating conditions obtained through the cross-correlation analysis, the input data and output data are matched, and the PCA dimensionality reduction method is used. The order after dimensionality reduction is used as the input of the data-driven model. Step 1.3.3: Build a data-driven model, namely a residual neural network; Step 1.4: Preprocess the data driven model to obtain the final digital twin mechanism model; Step 2: After obtaining real-time sensor data, construct velocity field, pressure field, and temperature field models and fuse them with the sensor data as input data for the data-driven model. Then, obtain the lag time within short-term operating conditions through cross-correlation analysis. After inputting the data into the data-driven model, obtain the amount of carbon emissions from the furnace top. Combined with the lag time, determine the time of the predicted carbon emissions, thereby realizing the prediction of carbon emission anomalies.

2. The method for predicting abnormal carbon emissions from a blast furnace based on digital twins according to claim 1 is characterized in that: The step 1 specifically utilizes digital twin technology to solve the Navier-Stokes NS equations of two-dimensional incompressible fluid; by solving the Navier-Stokes NS equations and the heat conduction differential equation, the velocity field, temperature field and pressure field are determined.

3. The method for predicting abnormal carbon emissions from a blast furnace based on digital twins according to claim 1 is characterized in that: Step 1.3.3 is specifically as follows: after obtaining the carbon lag time under the current operating conditions, deep learning is used to predict carbon emission anomalies; specifically, an optimized convolutional neural network (CNN) is used to diagnose carbon emission anomalies under various operating conditions; the optimized convolutional neural network (CNN) takes the feature input as the starting point, first through two sets of parallel convolution operations, each set contains 32 convolution kernels, and then accesses the ReLU activation function. The feature map enters the next stage after addition fusion, and then the feature map is reduced in size through the pooling layer, and the size is adjusted by padding with 0 to complete the preprocessing; in the next stage, the number of convolution layers is expanded to 64, and the activation function is continued to be used for processing. After fusion, the features enter the final fully connected layer to complete feature extraction and mapping; the entire network realizes layer-by-layer feature extraction and information aggregation through multi-level convolution, pooling and fully connected operations. The output vector y is calculated as follows: y=H(x)=F(x,{W i })+x (3) Where F(x, {W i }) represents the learned residual mapping, and x is the input vector.

4. The method for predicting abnormal carbon emissions from a blast furnace based on digital twins according to claim 1 is characterized in that: The preprocessing described in step 1.4 specifically normalizes the data and applies the normalization formula to scale the data to the range of [-1, 1]: x norm =2*(x-x min ) / (x max -x min )-1 (4) In the formula, x represents the input data that needs to be preprocessed, x norm represents the normalized result, and x min and x max Represent the minimum and maximum values ​​of a variable set, respectively.

Images