Blast furnace temperature anomaly prediction method based on digital twinning

Abstract

The application provides a blast furnace temperature anomaly prediction method based on digital twinning, relates to the technical field of digital twinning, determines an adaptive diffusion coefficient according to a tuyere raceway area image collected in real time, can avoid errors caused by an adaptive diffusion coefficient not being a fixed value in the prior art, and can determine an energy conservation equation based on the determined adaptive diffusion coefficient, and then solve a velocity field according to real-time observation data corresponding to the tuyere raceway area image, a mass conservation equation and a momentum conservation equation, input the velocity of the velocity field into the energy conservation equation corresponding to the tuyere raceway area image, solve the equation by using a finite difference method, obtain a temperature field, and then extract temperature field characteristics of the temperature field, process the temperature field characteristics by using an abnormal situation prediction model, and obtain temperature labels of the temperature field, and the experimental results show that the application improves the blast furnace temperature anomaly prediction and significantly improves real-time performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of digital twinning, in particular to a blast furnace temperature anomaly prediction method based on digital twinning. BACKGROUND

[0002] Ironmaking is a complex physicochemical process aimed at reducing a mixture of iron ore, coke, and flux in a blast furnace into molten iron. The ironmaking process has complex thermodynamic, fluid dynamic, and chemical reaction characteristics, requiring effective and precise control to stabilize production. The complexity of the ironmaking process is mainly reflected in temperature variation, gas flow distribution, reaction equilibrium, and equipment degradation. It is crucial to stabilize the ironmaking process to produce high-quality molten iron; in particular, it is necessary to monitor operating conditions and detect abnormal conditions.

[0003] Various monitoring methods have been used to detect abnormal operating conditions in blast furnaces. Existing methods are divided into model-based methods, knowledge-based methods, and data-driven methods. While model-based or knowledge-based methods show limitations in dealing with the complexity of the blast furnace ironmaking process, various data-driven methods have been explored. For example, deep weighted joint distributed adaptive network (DWJDAN) combines convolutional neural networks and transfer learning techniques to address the scarcity and inconsistency of real-time data. Adaptive dynamic explainable analysis of stationary subspace analysis (DiASSA) separates dynamic and static components through an inference observation decomposition and iterative modeling strategy. A framework that integrates principal component analysis (PCA) methods into a graph neural network (GNN) framework is proposed for diagnosing faults in blast furnaces. The multivariate statistical process control model (GTM-SPC) is improved, in which TOPSIS and grey models are used for optimization. The time series prediction model (SCPM) is based on spatio-temporal aggregation and further uses Granger causality analysis to detect faults in industrial processes. An unsupervised fault monitoring framework combines gated recurrent units (GRUs), Gaussian mixture models (GMMs), and autoencoders (AEs). The joint training method improves performance by minimizing maximum entropy through joint optimization of feature extractors and classifiers. The RSLADCGRU proposed by Li et al. uses graph neural networks to describe the sparse relationships of process variables. The local-DBKSSA method uses dynamic generalized nonlinear feature extraction and local statistics to monitor nonlinear and dynamic characteristics. Regularized mutual kernel analysis stationary subspace analysis (RMK-ASSA) is combined with a deep generalized stationary kernel network (DBSKNet) to handle the nonlinear and non-stationary characteristics of the data. Principal component analysis (PCA) and independent component analysis (ICA) are combined to monitor and diagnose abnormal conditions in datasets with Gaussian or non-Gaussian distributions. Multiple local manifold learning improves fault detection by exploiting local geometric structures. A deep learning method called DSKL-SVM processes non-stationary data by analyzing the steady-state subspace and deep kernel networks. A multi-channel dynamic graph convolutional network (MDGCN) is developed to represent the dynamic characteristics of industrial processes by assigning dynamic weights and applying multi-level feature learning. Multi-cooperative preserving projection (MCPP) is used to extract spatio-temporal structures; multi-source heterogeneous data are used to reduce the impact of industrial noise.

[0004] Despite some progress in developing various data-driven methods, existing methods face many technical obstacles: First, these methods heavily rely on the quality and sufficiency of data; however, it is extremely difficult to directly obtain sufficient data related to various faults. Second, the results of data-driven methods are often not supported by physical explanations; they are uninterpretable and opaque. Third, data-driven methods show limited robustness and generalization ability when data is non-stationary, dynamic, and complex. Finally, existing data-driven methods are not satisfactory in terms of accuracy in detecting new or rare faults. SUMMARY

[0005] In view of the deficiencies of the prior art, the purpose of the present application is to provide a blast furnace temperature anomaly prediction method based on digital twinning, comprising:

[0006] Step 1: Real-time acquisition of production data in the blast furnace smelting process, the production data comprising a plurality of tuyere raceway zone images and real-time observation data, the real-time observation data at least comprising speed data of the tuyere outlet and temperature data of the tuyere outlet, each tuyere raceway zone image corresponding to real-time observation data;

[0007] Step 2: For each tuyere raceway zone image, determine the adaptive diffusion coefficient corresponding to the tuyere raceway zone image, and determine the energy conservation equation corresponding to the tuyere raceway zone image according to the adaptive diffusion coefficient;

[0008] Step 3: For each tuyere raceway zone image, according to the real-time observation data corresponding to the tuyere raceway zone image, and the mass conservation equation and the momentum conservation equation, solve to obtain a velocity field, input the velocity of the velocity field into the energy conservation equation corresponding to the tuyere raceway zone image, and solve the equation by using the finite difference method to obtain a temperature field;

[0009] Step 4: Feature extraction is performed on the temperature field to obtain temperature field features, and an abnormal situation prediction model based on a random forest regression algorithm is used to process the temperature field features to obtain temperature labels of the temperature field, the temperature labels including abnormal low temperature, normal temperature and abnormal high temperature; the abnormal situation prediction model is obtained by model training based on a plurality of first training samples, the first training samples comprising first input samples and first output samples, the first input samples being historical temperature field features, and the first output samples being temperature labels corresponding to the historical temperature field features.

[0010] Optionally, in step 2, for each tuyere raceway zone image, the diffusion coefficient corresponding to the tuyere raceway zone image is determined, specifically comprising:

[0011] For each tuyere rotation zone image, a texture feature of the tuyere rotation zone image is extracted by a gray level co-occurrence matrix, and then a texture feature corresponding to each tuyere rotation zone image is obtained; the texture feature, the speed data of the air outlet and the temperature data of the air outlet are input into an adaptive diffusion coefficient construction model to obtain an adaptive diffusion coefficient corresponding to the tuyere rotation zone image, wherein the adaptive diffusion coefficient construction model is obtained by model training based on a plurality of second training samples, the second training samples include second input samples and second output samples, the second input samples are texture features of historical tuyere rotation zone images, speed data of air outlets and temperature data of air outlets, and the second output samples are adaptive diffusion coefficients of the historical tuyere rotation zone images.

[0012] Optionally, the texture feature includes contrast, correlation, energy and homogeneity.

[0013] Optionally, in step 3, the velocity field is obtained according to the real-time observation data corresponding to the tuyere rotation zone image, and the mass conservation equation and the momentum conservation equation, and specifically includes:

[0014] The real-time observation data corresponding to the tuyere rotation zone image is input into the mass conservation equation and the momentum conservation equation to obtain a target equation, the target equation is discretized by a finite volume method, an initial iteration number is set, an initial velocity field under the current iteration number is set, and a predicted velocity field is obtained according to the initial velocity field and the discretized target equation, and the predicted velocity field is revised to obtain a revised velocity field. It is judged whether the current iteration number reaches a preset threshold, in the case that the current iteration number reaches the preset threshold, the revised velocity field is output, in the case that the current iteration number does not reach the preset threshold, the current iteration number is increased by one and taken as a new current iteration number, and the setting of the initial velocity field under the current iteration number is returned to be executed until the current iteration number reaches the preset threshold.

[0015] Optionally, the temperature field feature includes the average temperature of the tuyere rotation zone, the average temperature gradient, the maximum temperature gradient, the maximum temperature difference, the standard deviation of the temperature gradient and the median temperature gradient.

[0016] Optionally, after the velocity field is obtained in step 3, it further includes:

[0017] According to the size and shape of the blast furnace, an elliptic partial differential equation is used to generate a body fitting coordinate grid, the velocity field is combined with the body fitting coordinate grid to obtain a visual representation of the velocity field.

[0018] Optionally, after the temperature field is obtained in step 3, it further includes:

[0019] Based on the size and shape of the blast furnace, a volume-fitted coordinate grid is generated using elliptic partial differential equations. The temperature field is then combined with the volume-fitted coordinate grid to obtain a visual representation of the velocity field.

[0020] The beneficial effects of adopting the above technical solution are as follows:

[0021] This invention determines an adaptive diffusion coefficient based on real-time acquired images of the tuyere swirl zone, avoiding errors caused by the non-fixed value of the adaptive diffusion coefficient in existing technologies. Based on the determined adaptive diffusion coefficient, the energy conservation equation can be determined. Then, using real-time observation data corresponding to the tuyere swirl zone image, along with the mass and momentum conservation equations, the velocity field is obtained. The velocity of this velocity field is input into the energy conservation equation corresponding to the tuyere swirl zone image, and the finite difference method is used to solve the equation, yielding the temperature field. The temperature field features are then extracted and processed using an anomaly prediction model to obtain temperature labels. Experimental results show that this invention improves the prediction of blast furnace temperature anomalies and significantly enhances real-time performance. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a method for predicting blast furnace temperature anomalies based on digital twins, as described in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of a process for constructing a model using an adaptive diffusion coefficient, as described in an embodiment of the present invention.

[0024] Figure 3 This is a flowchart illustrating a method for solving a mechanism model in an embodiment of the present invention;

[0025] Figure 4 These are images of the air vent swirl zone under three temperature conditions in the embodiments of the present invention. Among them, Figure (a) is an image of the air vent swirl zone at an abnormally low temperature, Figure (b) is an image of the air vent swirl zone at a normal temperature, and Figure (c) is an image of the air vent swirl zone at an abnormally high temperature.

[0026] Figure 5 Figure (a) shows the temperature field images under three temperature conditions in the embodiments of the present invention, wherein Figure (b) is the temperature field image of abnormally low temperature, Figure (c) is the temperature field image of normal temperature, and Figure (d) is the temperature field image of abnormally high temperature.

[0027] Figure 6 This is a schematic diagram of a process for processing abnormal situations using an abnormal situation prediction model in an embodiment of the present invention;

[0028] Figure 7 This is a schematic diagram showing the comparison results of the diffusion coefficients of the random forest model for measurement and prediction in an embodiment of the present invention;

[0029] Figure 8 This is a schematic diagram showing the comparison results of the average temperature at the top edge under different diffusion coefficients in an embodiment of the present invention;

[0030] Figure 9 This is a schematic diagram showing the comparison results of the learning curves of the four algorithms in this embodiment of the invention;

[0031] Figure 10 This is a schematic diagram of the confusion matrix in an embodiment of the present invention. Detailed Implementation

[0032] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0033] To address the problems existing in current technologies, this invention improves existing data-driven methods for monitoring the ironmaking process in blast furnaces through two innovations: (1) A digital twin (DT) framework based on hybrid modeling is proposed, which combines mechanistic modeling and data-driven modeling to detect anomalies in blast furnaces. (2) An adaptive diffusion coefficient is introduced to improve the accuracy of the temperature determined by simplifying the energy conservation equation. Specifically, this invention provides a blast furnace temperature anomaly prediction method based on digital twins, combining... Figure 1 This may include the following steps:

[0034] Step 1: Real-time acquisition of production data during the blast furnace production process. The production data includes multiple tuyere swirl zone images and real-time observation data. The real-time observation data includes at least the velocity data and temperature data of the tuyere outlet. Each tuyere swirl zone image corresponds one-to-one with the real-time observation data.

[0035] Step 2: For each vortex image, determine the adaptive diffusion coefficient corresponding to the vortex image, and determine the energy conservation equation corresponding to the vortex image based on the adaptive diffusion coefficient.

[0036] It should be noted that, traditionally, the diffusion coefficient α in the energy conservation equation is fixed, as shown in the following formula:

[0037]

[0038] Where k represents thermal conductivity, ρ is the density of the material, and c v This indicates the constant-volume heat capacity.

[0039] In traditional methods, the diffusion coefficient α is treated as a constant, simplifying the modeling of the heat transfer process. However, this assumption may not accurately reflect the complexity of real-world systems. For example, in blast furnace operation, heat transfer is affected by dynamic factors such as changes in material composition, furnace operating conditions, and temperature gradients. A fixed diffusion coefficient assumes homogeneous material properties, but in reality, due to phase transitions, chemical reactions, and non-uniform material distribution, thermal conductivity k and specific heat capacity c vary significantly. v Significant variations can occur. These variations can lead to biases in temperature predictions. Specifically, the texture features of flame images (contrast, correlation, brightness, and uniformity) are closely related to the diffusion coefficient. Higher contrast indicates more significant differences, and intensity variations in the flame typically increase gas mixing and the diffusion coefficient. High brightness is associated with intense combustion reactions, leading to an increased diffusion coefficient. Higher flame correlation indicates more stable gas flow, thus reducing the diffusion coefficient, while lower uniformity reflects unstable gas flow, which increases the diffusion coefficient. Furthermore, the velocity and temperature at the nozzle directly affect the diffusion coefficient; higher nozzle velocities and temperatures enhance gas energy and mixing, thereby increasing the diffusion coefficient.

[0040] To address these issues, an adaptive diffusion coefficient model can be employed. This model adjusts the diffusion coefficient based on real-time data, providing a more accurate representation of heat transfer kinetics. By considering variability and adapting to changing conditions, the adaptive model can significantly improve the accuracy of temperature assessments and the overall system performance.

[0041] In step 2, for each vent swirl zone image, the diffusion coefficient corresponding to the vent swirl zone image is determined, combined with... Figure 2 Specifically, it includes:

[0042] For each image of the wind swirl zone, texture features of the wind swirl zone image are extracted using the gray-level co-occurrence matrix, i.e. Figure 2 The texture features extracted by the wind vent image texture analysis and feature extraction module include contrast, correlation, energy, and homogeneity; thereby obtaining the texture features corresponding to each wind vent swirling zone image; the texture features, air outlet velocity data, and air outlet temperature data are input into the adaptive diffusion coefficient construction model to obtain the adaptive diffusion coefficient corresponding to the wind vent swirling zone image, wherein the adaptive diffusion coefficient construction model is obtained by training the model based on multiple second training samples, the second training samples include second input samples and second output samples, the second input samples are the texture features of historical wind vent swirling zone images, air outlet velocity data, and air outlet temperature data, and the second output samples are the adaptive diffusion coefficients of the historical wind vent swirling zone images;

[0043] The air outlet temperature data is obtained visually, while the air outlet speed data is collected by sensors.

[0044] Based on this, the diffusion coefficient is determined, and then the diffusion coefficient is substituted into the energy conservation equation to obtain the final energy conservation equation.

[0045] Step 3: For each image of the swirling zone of the swirling zone, based on the real-time observation data corresponding to the image of the swirling zone of the swirling zone, as well as the mass conservation equation and momentum conservation equation, solve to obtain the velocity field. Input the velocity of the velocity field into the energy conservation equation corresponding to the image of the swirling zone of the swirling zone of the swirling zone, and use the finite difference method to solve the equation to obtain the temperature field.

[0046] Combination Figure 3 This is a flowchart illustrating the entire process of step 3. The mechanistic model includes the energy conservation equation, the mass conservation equation, and the momentum conservation equation. In practical implementation, the energy conservation equation needs to be simplified to obtain the simplified energy conservation equation, i.e. Figure 3 The gas temperature model in the equation can be expressed by the following formula:

[0047]

[0048] The mass conservation equation is expressed by the following formula:

[0049]

[0050] The momentum conservation equation is expressed by the following formula:

[0051]

[0052] Combination Figure 3 In step 3, the velocity field is obtained by solving the real-time observation data corresponding to the image of the wind vortex area, as well as the mass conservation equation and momentum conservation equation. Specifically, this includes:

[0053] Real-time observation data corresponding to the wind vortex area image is input into the mass conservation equation and momentum conservation equation to obtain the objective equation. The objective equation is discretized using the finite volume method. An initial number of iterations is set, and an initial velocity field is set for the current iteration number. Based on the initial velocity field and the discretized objective equation, the predicted velocity field is solved and revised to obtain the revised velocity field. It is then determined whether the current iteration number has reached a preset threshold. If the current iteration number has reached the preset threshold, the revised velocity field is output. If the current iteration number has not reached the preset threshold, the current iteration number is incremented by one and used as the new current iteration number. The process continues until the current iteration number reaches the preset threshold. Figure 3 The method for solving flow gas models.

[0054] After obtaining the velocity field in step 3, the following steps are also included:

[0055] Based on the size and shape of the blast furnace, a volume-fitting coordinate grid is generated using elliptic partial differential equations. The velocity field is then combined with this grid to obtain a visual representation of the velocity field. Similarly, based on the size and shape of the blast furnace, a volume-fitting coordinate grid is generated using elliptic partial differential equations. The temperature field is then combined with this grid to obtain a visual representation of the velocity field. Figure 3 Medium mesh generation section, The equation is an elliptic partial differential equation, where 29.8 meters is the height of the blast furnace, and 273 horizontal lines and 61 vertical lines represent the grid dimensions.

[0056] Combination Figure 3 Step 3 may specifically include the following parts:

[0057] A. Mesh Generation and Boundary Conditions

[0058] The blast furnace consists of five main parts: throat, body, waist, belly, and bottom. A coordinate (BFC) mesh adapted to the body was created using elliptic partial differential equations. To reduce computational complexity, the solid model was simplified using the symmetric structure of the nozzles and furnace lining, resulting in a two-dimensional model. The modeling region extends from the bottom of the nozzles to the furnace throat, where lateral temperature measurements were performed. The following boundary conditions (BCs) were defined to simulate the temperature distribution within the blast furnace:

[0059] (1) Inlet: The nozzle is modeled as an inlet boundary, where the values ​​of the field variables are set based on field measurements.

[0060] (2) Outlet: The top of the furnace is modeled as an outlet, and its pressure is specified by measurement.

[0061] (3) No slip surface: The inner wall of the furnace is designed as a no slip surface for the flow of molten metal; therefore, the wall is impermeable and the relative velocity is zero.

[0062] (4) Symmetry plane: The magnitude and gradient of the velocity perpendicular to the symmetry plane are set to 0.

[0063] B. Gas Flow Model

[0064] Solving the Navier-Stokes equations requires modeling the coupling of pressure and velocity in the gas flow. In the flow model, the source term F typically represents the drag effect assessed by the Ergenche equations. Furthermore, a lift source term F' is introduced to account for the buoyancy effect caused by the temperature gradient. The velocity field can be obtained by simultaneously solving the mass and momentum conservation equations.

[0065] C. Gas Temperature Model

[0066] The unsimplified energy conservation equation encompasses various energy transfer mechanisms, such as convection, heat conduction, thermal radiation, work transfer, chemical energy changes, heat sources, and latent heat of phase transitions. In the gas temperature model, the temperature distribution reflects the relationships between gas convection, heat conduction, heat exchange between the gas and solid phases, and chemical reactions, and its temperature distribution is characterized by spatiotemporal variations. To account for the influence of chemical reactions on the temperature distribution, a source term Q for the chemical reaction is specified.

[0067] In the ironmaking process, the temperature distribution within the blast furnace is a crucial indicator of its operational status. Under normal circumstances, temperature changes are stable, resulting in a smooth smelting process. However, due to the numerous factors affecting the actual ironmaking process, temperature changes can become unstable under abnormal conditions. This invention studies two typical abnormal situations: abnormal decreases and increases in blast furnace temperature. A decrease in blast furnace temperature affects the efficiency and quality of the smelting process, potentially leading to overload of the cooling system. An increase in blast furnace temperature accelerates the loss of refractory material in the furnace wall, affecting the furnace lining and causing blast furnace operation interruption. Therefore, it is necessary to predict abnormal temperature conditions, which is specifically achieved through step 4.

[0068] like Figure 4 Figure (a) shows the image of the air swirl zone at abnormally low temperatures, Figure (b) shows the image of the air swirl zone at normal temperatures, and Figure (c) shows the image of the air swirl zone at abnormally high temperatures. (Combined...) Figure 5 Figure (a) shows the temperature field image at abnormally low temperatures, Figure (b) shows the temperature field image at normal temperatures, and Figure (c) shows the temperature field image at abnormally high temperatures. Compare these figures. Figure 4 and Figure 5 The results show a significant difference between the measured and predicted temperature fields. These differences indicate variations in heat transfer conditions within the blast furnace and underscore the necessity of extracting and characterizing the temperature distribution, thus necessitating step 4.

[0069] Step 4: Combining Figure 6 The temperature field is feature extracted to obtain temperature field features, which include the average temperature, average temperature gradient, maximum temperature gradient, maximum temperature difference, standard deviation of temperature gradient, and median temperature gradient in the wind vortex zone. Figure 1 and Figure 6 The furnace body temperature field is the same as the temperature field in step 4. Figure 6 The region of interest is the wind vortex area.

[0070] An anomaly prediction model based on random forest regression algorithm is used to process temperature field features to obtain temperature labels for the temperature field. The temperature labels include abnormally low temperature, normal temperature, and abnormally high temperature. The anomaly prediction model is trained based on multiple first training samples. The first training samples include first input samples and first output samples. The first input samples are historical temperature field features, and the first output samples are the temperature labels corresponding to the historical temperature field features.

[0071] During model training, an initial model is first constructed based on the random forest regression algorithm, and then the initial model is trained. The initial model can refer to the model in the literature "Random Forests".

[0072] The temperature labels, including abnormally low temperature, normal temperature, and abnormally high temperature, are manually marked. Abnormally low temperature indicates that the temperature in the air vent swirl zone is lower than a first preset threshold, which can be 1710℃. Normal temperature indicates that the temperature in the air vent swirl zone is within a threshold range, which can be 1726℃ to 2000℃. Abnormally high temperature indicates that the temperature in the air vent swirl zone is higher than a second preset threshold, which can be 2234℃ to 2482℃.

[0073] Experiments were conducted to verify the above content, including:

[0074] A. Evaluation of the adaptive diffusion coefficient

[0075] To evaluate the applicability of machine learning regression models to specific regression tasks, a comparative study was conducted, testing 1000 constructed samples using Support Vector Machines (SVM), K-Nearest Neighbors (KNN), Random Forest (RF), Linear Regression (LR), Ridge Regression (RR), and Gradient Boosting Regression (GBR). 5% of the samples were used as the test dataset. Root Mean Square Error (RMSE) and Mean Absolute Percentage Error (MAPE) were used as evaluation metrics. These two metrics were calculated as follows:

[0076]

[0077] Among them, y i and These represent the measured value and the predicted value, respectively; n is the sample size.

[0078] Table 1 compares the results of the six regression models from the perspective of two evaluation criteria. The comparison results show that the RF regression model predicts the best diffusion coefficient. Figure 7 The measured values ​​were further compared with the diffusion coefficients predicted by the RF regression model.

[0079] Table 1 Comparison of different regression models

[0080] Regression model RMSE(10 -3 m 2 / s)]]> MAPE (%) SVM 5.497 4.413 KNM 2.102 1.480 LR 2.373 1.639 RR 2.354 1.626 GBR 2.156 1.460 RF 2.040 1.438

[0081] B. Temperature distribution

[0082] To verify the effectiveness of using an adaptive diffusion coefficient to improve the accuracy of predicted temperature fields, the results using the adaptive diffusion coefficient were compared with those using a fixed diffusion coefficient.

[0083] In addition, such as Figure 8 As shown, the temperature at the top edge of the furnace was measured and averaged. Clearly, when the diffusion coefficient is adaptive, the average error is 3.17%; in comparison, the average error is 6.90% when a fixed diffusion coefficient is used.

[0084] C. Evaluation of the abnormal situation prediction model

[0085] To evaluate the accuracy of the proposed radio frequency spectrum (RF) method, it was compared with three other classification models: K-Nearest Neighbors (KNN), Naive Bayes (NB), and Support Vector Machine (SVM). 30% of the samples were used as test samples, and Table 2 compares the results of these four classifiers from the perspective of classification accuracy.

[0086] Table 2 Comparison of different classification methods

[0087] Method Accuracy (%) NB 83.77 SVM 87.35 KNM 95.58 RF 97.49

[0088] To evaluate the generalization ability of the proposed model, 5-fold cross-validation was performed following these steps: The training dataset was randomly divided into 5 subsets. One subset was used as the validation set, and the other 4 subsets were used as the training dataset. This cycle was repeated 5 times, with 5 sets of results.

[0089] The training and validation results were finally obtained. The performance evaluation results of one classification method are as follows:

[0090]

[0091] Where, N correct N represents the number of correctly classified samples; total It represents the total number of data samples.

[0092] The average of these 5 groups is considered as the final training score and the model's validation score. Figure 9 The learning curve for each model is shown.

[0093] Table 2 and Figure 9 The results show that Random Forest (RF) performs best. The overall performance of the constructed radio frequency spectrum classification model is evaluated using accuracy, precision, and recall. Table 3 presents the specific results. Figure 10A confusion matrix is ​​presented, visually demonstrating the model's classification performance across different categories. The model achieves an accuracy of 97.49% on the test set. The proposed prediction method demonstrates excellent accuracy and a significant improvement in prediction speed. To monitor and detect anomalies, all these methods are implemented in Python 3.8 and executed on a local computer equipped with an Intel Core i5-13600KF CPU (5.1 GHz) and 64 GB of memory, with prediction times consistently under 5 minutes, thus meeting the requirements of real-time digital twin applications.

[0094] Monitoring and detecting anomalies is crucial for ensuring safety and improving production efficiency. We propose a hybrid modeling-based digital twin (DT) framework for detecting anomalies in production. First, a simplified mechanistic model is established, and an adaptive diffusion coefficient is introduced to improve accuracy and determine the temperature field. Then, the temperature field is described by several key attributes, which are used as inputs to a classification model for training and anomaly identification. Experimental results show that the proposed DT framework effectively integrates mechanistic modeling and data-driven modeling. This integration significantly improves real-time performance.

[0095] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for predicting blast furnace temperature anomalies based on digital twins, characterized in that, include: Step 1: Real-time acquisition of production data during the blast furnace production process. The production data includes multiple tuyere swirl zone images and real-time observation data. The real-time observation data includes at least the velocity data and temperature data of the tuyere outlet. Each tuyere swirl zone image corresponds one-to-one with the real-time observation data. Step 2: For each vortex image, determine the adaptive diffusion coefficient corresponding to the vortex image, and determine the energy conservation equation corresponding to the vortex image based on the adaptive diffusion coefficient. Step 3: For each image of the swirling zone of the swirling zone, based on the real-time observation data corresponding to the image of the swirling zone of the swirling zone, as well as the mass conservation equation and momentum conservation equation, solve to obtain the velocity field. Input the velocity of the velocity field into the energy conservation equation corresponding to the image of the swirling zone of the swirling zone of the swirling zone, and use the finite difference method to solve the equation to obtain the temperature field. Step 4: Extract features from the temperature field to obtain temperature field features. Process the temperature field features using an anomaly prediction model based on the random forest regression algorithm to obtain temperature labels for the temperature field. The temperature labels include abnormally low temperature, normal temperature, and abnormally high temperature. The anomaly prediction model is trained based on multiple first training samples. The first training samples include first input samples and first output samples. The first input samples are historical temperature field features, and the first output samples are the temperature labels corresponding to the historical temperature field features. In step 2, for each vent swirl zone image, the diffusion coefficient corresponding to the vent swirl zone image is determined, specifically including: For each vent swirl zone image, texture features are extracted using a gray-level co-occurrence matrix to obtain the texture features corresponding to each vent swirl zone image. The texture features, air outlet speed data, and air outlet temperature data are input into an adaptive diffusion coefficient construction model to obtain the adaptive diffusion coefficient corresponding to the vent swirl zone image. The adaptive diffusion coefficient construction model is obtained by training the model based on multiple second training samples. The second training samples include second input samples and second output samples. The second input samples are the texture features of the historical vent swirl zone image, the air outlet speed data, and the air outlet temperature data. The second output sample is the adaptive diffusion coefficient of the historical vent swirl zone image. In step 3, the velocity field is obtained by solving the real-time observation data corresponding to the image of the wind vortex area, as well as the mass conservation equation and the momentum conservation equation. Specifically, this includes: Real-time observation data corresponding to the image of the wind vortex zone is input into the mass conservation equation and momentum conservation equation to obtain the objective equation. The objective equation is discretized using the finite volume method. An initial number of iterations is set, and an initial velocity field is set for the current number of iterations. Based on the initial velocity field and the discretized objective equation, the predicted velocity field is solved and revised to obtain the revised velocity field. It is then determined whether the current number of iterations has reached a preset threshold. If the current number of iterations has reached the preset threshold, the revised velocity field is output. If the current number of iterations has not reached the preset threshold, the current number of iterations is incremented by one and used as the new current number of iterations. The process then returns to execute: setting the initial velocity field for the current number of iterations until the current number of iterations reaches the preset threshold.

2. The method for predicting blast furnace temperature anomalies based on digital twins according to claim 1, characterized in that, The texture features include contrast, correlation, energy, and homogeneity.

3. The method for predicting blast furnace temperature anomalies based on digital twins according to claim 1, characterized in that, The temperature field characteristics include the average temperature, average temperature gradient, maximum temperature gradient, maximum temperature difference, standard deviation of temperature gradient, and median temperature gradient of the wind vortex zone.

4. The method for predicting blast furnace temperature anomalies based on digital twins according to claim 1, characterized in that, After obtaining the velocity field in step 3, the following steps are also included: Based on the size and shape of the blast furnace, a volume-fitted coordinate grid is generated using elliptic partial differential equations. The velocity field is then combined with the volume-fitted coordinate grid to obtain a visual representation of the velocity field.

5. The method for predicting blast furnace temperature anomalies based on digital twins according to claim 1, characterized in that, After obtaining the temperature field in step 3, the following steps are also included: Based on the size and shape of the blast furnace, a volume-fitted coordinate grid is generated using elliptic partial differential equations. The temperature field is then combined with the volume-fitted coordinate grid to obtain a visual representation of the velocity field.

Images