A method and system for analyzing converter slag state and predicting splashes

Abstract

The application discloses a method and system for analyzing converter slag state and predicting splashing, the method comprises the following steps: 1, constructing a multi-physical field simulation geometric model coupled with fluid mechanics and radiation transfer equation, discretizing the upper surface of the molten pool and setting a dynamic gas mass flow boundary condition; 2, running the simulation and extracting flame characteristic quantitative indicators; 3, collecting and processing industrial field flame images; 4, based on the feature processing and uncertainty modeling of the chaotic polynomial expansion, constructing a chaotic polynomial proxy model for the extracted flame image features; 5, establishing a permeability state recognition model and a splashing\return dry prediction. The application effectively processes the random uncertainty of the parameters in the flame image by using the chaotic polynomial method, improves the robustness of feature extraction and state inversion, realizes the quantitative recognition of the non-uniformity of the slag state space and the accurate prediction of the splashing\return dry risk.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a method and splash prediction system for analyzing converter slag state and predicting splashing based on multiphysics simulation and chaotic polynomial expansion. Background Technology

[0002] Converter steelmaking is a crucial process in steel production, characterized by rapid and intense molten pool reactions under extreme conditions such as high temperature, high pressure, and strong turbulence. During the blowing process, the state of the molten slag on the surface of the molten pool, especially its permeability—that is, its ability to allow gas to pass through—directly determines the gas release pattern within the furnace, thus influencing the morphology and evolution of the furnace mouth flame. The furnace mouth flame, as a "barometer" of the molten pool reaction state, contains a wealth of process information. Traditionally, operators rely on visual observation of the flame's color, shape, and frequency of movement to determine the slag state (such as splashing and drying). This method is highly dependent on personal experience, subjective, and difficult to monitor in real time.

[0003] With the improvement of machine vision technology and industrial camera resolution, converter monitoring technology based on flame image analysis has gradually become a research hotspot. However, although the number of flame images collected in industrial sites is enormous, their features often exhibit randomness and uncertainty due to sensor noise, furnace mouth smoke obstruction, and dynamic changes in operating conditions. Furthermore, in industrial practice, obtaining high-quality labeled data with precise process tags is difficult, and often only a small sample of labeled data is available. In addition, flame image features (such as brightness and pulsation frequency) are inherently random, and existing methods lack effective means to systematically characterize this random uncertainty, resulting in prediction results lacking confidence support and making it difficult to achieve highly reliable safety early warnings.

[0004] In recent years, the Polynomial Chaos Expansion (PCE) method, first proposed by Wiener, is based on the idea of ​​representing the system response as a weighted sum of a series of orthogonal polynomial basis functions. This allows for efficient handling of the randomness of input variables with relatively low computational cost, and the acquisition of the probability distribution characteristics of the output response. In traditional applications, PCE is primarily used in structural reliability analysis and computational fluid dynamics verification, employing surrogate models to assess the impact of input parameter perturbations on simulation results. However, in the fields of image processing and feature analysis, especially in complex industrial scenarios such as converter steelmaking with high temperature, high dust, and strong interference, the use of PCE to systematically model and analyze the random uncertainties of flame image features to improve the robustness of slag state inversion has not yet been reported.

[0005] In existing technologies, flame images from industrial sites are typically acquired directly as data sources to train models for determining slag states. However, acquiring measured flame image data from industrial sites is inherently difficult: firstly, image quality is hard to guarantee due to harsh conditions such as high temperatures, smoke, and varying lighting, resulting in insufficient sample sizes and significant random noise in the measured data. Existing methods based on measured data struggle to effectively handle this uncertainty, limiting prediction accuracy and robustness. Secondly, measured data only reflects the final appearance of the slag state and cannot reveal the underlying physical evolution mechanisms. Although constructing simulation models that couple fluid dynamics and radiative transfer can mechanistically simulate the flame evolution process under different slag permeability states and generate theoretical data covering various conditions, there are currently no applications that use such high-fidelity simulation models as an effective substitute or supplement to measured data for chaotic polynomial expansion modeling to improve state inversion accuracy.

[0006] Therefore, how to integrate simulation data and measured data to build a technical solution that can simultaneously overcome the difficulties in data acquisition, quantify random uncertainty, and enhance the physical interpretability of the model is an urgent problem to be solved in the field of converter steelmaking process monitoring.

[0007] Defects and shortcomings of existing technology: Although converter steelmaking process monitoring technology has been researched and developed for many years, and various methods have attempted to solve the problems of splash prediction and slag state identification from different perspectives, existing technologies still suffer from limitations due to the extreme operating conditions of converter steelmaking, such as high temperature, high dust levels, and strong interference. These limitations include a weak correlation between the acquisition method of the monitoring signal and the actual state of the molten pool, difficulty in effectively removing environmental noise and random interference during signal processing, and the fact that most methods can only reflect a single dimension of process information, failing to comprehensively characterize the complex evolution of the slag state. The shortcomings and deficiencies of various mainstream methods are as follows:

[0008] 0. Deficiencies of Traditional Manual Observation Methods: Traditional manual observation methods rely primarily on the visual experience of operators, judging the state of molten slag by observing the color, shape, and frequency of the flame at the furnace opening and listening to the blowing sounds. However, this method is limited by factors such as individual experience and skill level, making it difficult to guarantee the stability and accuracy of the detection results. Different workers' judgments may vary significantly, exhibiting strong subjectivity and poor real-time performance, often only being detected after splashing occurs, making continuous monitoring and early warning difficult. Furthermore, the high rate of missed detections is influenced by factors such as worker condition and fatigue, failing to meet the precision control requirements of modern steelmaking.

[0009] 1. Limitations of Furnace Gas Analysis Methods: Furnace gas analysis methods assess the intensity of carbon-oxygen reactions within the furnace and predict splashing risks by monitoring changes in the content of components such as CO, CO2, and O2 in the flue gas at the furnace in real time. However, this method suffers from significant signal lag—there is a time delay between the generation of flue gas in the molten pool and the appearance of detection results, making it difficult to provide immediate warnings. Furthermore, air intake at the furnace inlet dilutes the flue gas components, interfering with measurement accuracy. More importantly, this method only reflects carbon-oxygen reaction information and cannot comprehensively assess the slag foaming state and the actual operating conditions of the molten pool. This single-dimensional information is insufficient to accurately predict splashing accidents caused by complex, multi-factor coupling.

[0010] 2. Limitations of Audio Analysis Methods: Audio analysis collects sound signals from the blowing process and analyzes the relationship between changes in audio intensity and frequency and the height of the slag foam layer. However, this method has a significant monitoring blind spot—when slag foaming is severe, the audio signal is prone to saturation or distortion, failing to accurately reflect the actual furnace conditions. Furthermore, significant environmental noise interference and complex signal processing further complicate the process, requiring improvement in accuracy.

[0011] 3. Limitations of the Oxygen Lance Vibration Monitoring Method: The oxygen lance vibration monitoring method monitors the vibration acceleration data of the oxygen lance in the X, Y, and Z axes, calculates the vibration frequency of the molten slag foam based on the vibration characteristics, and then determines the probability of splashing risk. However, this technology is still in the development stage, and its application verification is insufficient. The physical correlation mechanism between vibration signals and splashing needs further clarification. Although triaxial vibration data can reflect the furnace conditions more comprehensively than unidirectional parameters, accurately separating splash-related feature information from complex vibration signals still faces technical challenges in signal processing and feature extraction.

[0012] 4. Limitations of AI-based methods: In real-world scenarios, flame images are often severely affected by smoke and dust, resulting in low image clarity and significant random noise. This uncertainty severely restricts the predictive performance of the model. While existing flame image-based methods can extract rich information about the furnace's internal state, they cannot effectively handle the random uncertainties in image features, thus limiting prediction accuracy. Summary of the Invention

[0013] This invention avoids interference from close-range sonar signals and data lag in furnace gas analysis by employing remote image acquisition to obtain image signals. This invention provides the following technical means to solve the problems existing in the background technology: 1. A multi-physics coupled simulation model with a clear mechanism was constructed to reflect the influence of dynamic changes in slag permeability on the flame generation process in the furnace window.

[0014] 2. The permeability state of the molten slag is parameterized and quantified, and then incorporated into the simulation model as a key boundary condition.

[0015] 3. By utilizing the chaotic polynomial expansion method, the random uncertainty in the flame image acquisition process can be effectively handled, thereby improving the robustness of feature extraction and state inversion.

[0016] 4. Establish a systematic verification framework, and use large-scale flame image data collected from industrial sites to quantitatively verify the accuracy of the simulation model in responding to different slag permeability states in terms of flow field structure and flame morphology, forming a closed loop of physical mechanism-numerical simulation-industrial data.

[0017] 5. Through innovation in models and methods, the non-uniformity of the spatial distribution of slag on the surface of the molten pool can be simulated and identified, thereby more accurately explaining and predicting complex flame observation phenomena.

[0018] The purpose of this invention is to provide a method and system for analyzing the state of converter slag and predicting splashing. By using a simulation model to replace some of the measured data, the difficulty of obtaining data is reduced and the prediction results are improved. Furthermore, the online identification of slag state and splashing prediction are achieved by using a constructed chaotic polynomial, thereby providing an interpretable, highly reliable method for online identification of slag state and splashing prediction that can accurately locate abnormal areas.

[0019] To achieve the above objectives, the present invention provides: A method for analyzing converter slag condition and predicting splashing includes: Step 1. Construct a multiphysics simulation geometric model that couples fluid dynamics and radiative transfer equations; Step 2. Run the simulation and extract the quantification indicators of flame characteristics: Run coupled simulation calculations, extract simulation results for the furnace window region, including velocity streamline contour maps, and extract the mean curvature from the velocity streamlines. As a core quantitative indicator.

[0020] Step 3. Acquiring and processing flame images from the industrial site: A high-temperature resistant industrial camera is installed on the side of the corresponding converter window to continuously capture flame images of the entire blowing process at a preset frequency, forming a real-shot image library.

[0021] Step 4. Feature processing and uncertainty modeling based on chaotic polynomial expansion: Multi-dimensional feature extraction is performed on the acquired flame images: texture feature extraction, color feature extraction, dynamic feature extraction, and key physical feature extraction; A random forest recursive elimination method is used for feature selection to determine the input variables; Construct a chaotic multinomial surrogate model, perform marginal distribution analysis on the selected features, determine the probability distribution type that each feature follows, and solve the chaotic multinomial coefficients using the training dataset through least squares regression. The training set is composed of the simulation data from step 1 and the measured labeled data from step 3, and is used as the training data for the chaotic polynomial surrogate model to complete the training of the chaotic polynomial surrogate model.

[0022] Step 5. Establish a permeability state identification model and splash prediction: Obtain multiple sets of permeability parameter combinations →Simulation features → Actual characteristics A mapping database is used to calibrate the simulation model; Establish a calculation based on real-shot flame images By combining the probability distribution output by the chaotic polynomial model, the combination of permeability parameters for each region of the current molten pool is inferred in reverse. The recognition model; Using a chaotic polynomial surrogate model as a splash / drying predictor, in converter steelmaking, when the obtained permeability parameters are consistently low and fluctuate drastically, or when the splash / drying probability output by the chaotic polynomial surrogate model exceeds the threshold, the system issues a warning that the slag permeability is deteriorating and there is a risk of splash / drying.

[0023] Step 6. Establish feedback and data augmentation mechanisms: To determine the accuracy of the prediction, when the system's output of splash / re-drying risk warnings does not match the actual working conditions, a feedback mechanism is established to augment the model with data; the augmented data is used to train and optimize the chaotic multinomial surrogate model to improve the model's accuracy.

[0024] Furthermore, in step 1, to establish a converter geometric model that includes the furnace body, the surface of the molten pool, and the furnace window outlet, the upper surface of the molten pool must be discretized into multiple independent control regions, and a parameterized dynamic gas mass inlet boundary condition must be set for each region i.

[0025] Furthermore, the multiphysics geometric model also includes the carbon-oxygen reaction kinetic field and the slag characteristic field.

[0026] Furthermore, in step 2, when there are N independent sector regions, the dynamic gas mass flow rate function is: in: For the steady-state basic gas generation rate, The transparency amplitude coefficient, This is a custom pulse width modulation function. It is a smooth start function.

[0027] Furthermore, step 2 is divided into ring grids, rectangular grids, or irregular partitions based on prior knowledge; the dynamic gas mass flow function adopts a piecewise function, a function based on stochastic processes, or a function coupled with other process parameters in real time.

[0028] More preferably, the camera used to acquire images in step 3 has more than 8 megapixels and the acquisition frequency is not less than 20 frames per second.

[0029] Furthermore, in step 4, the chaotic polynomial expansion transforms the system performance response function. Expand as , It is an orthogonal polynomial basis. For expansion coefficients, To truncate the order, orthogonal polynomials are used to expand the random features to quantify the uncertainty.

[0030] Furthermore, step 4, feature extraction, includes: Texture feature extraction: Extract entropy, energy, contrast and inverse difference matrix based on gray-level co-occurrence matrix; Color feature extraction: Extract first, second and third moments of H, S and V in HSV space; Dynamic feature extraction: Extract optical flow vector between consecutive frames based on optical flow method and calculate average optical flow velocity.

[0031] Furthermore, step 4 uses the Sobol index to analyze feature importance and screen out the key features that contribute the most to the prediction of splashing / drying.

[0032] Furthermore, if a splashing / drying event actually occurs, but the system fails to issue a warning before the event occurs or the warning probability is lower than the set threshold, then the flame image data segment is marked as a missed sample and added to the training set.

[0033] A converter slag state correlation analysis system includes a processor and a memory storing a computer program. When the processor executes the computer program, it implements the steps of the above-described method for analyzing converter slag state and predicting splashing.

[0034] The beneficial effects of this invention are: (1) It fundamentally improves the accuracy and physical interpretability of state recognition, transforming the traditional “black box” experience judgment into a quantitative diagnosis with a clear physical causal chain (permeability → airflow → streamline curvature), and the results are reliable and credible.

[0035] (2) Online quantitative identification of the non-uniformity of the molten slag state space is realized, which can accurately locate the local abnormal area on the surface of the molten pool, providing a direct basis for targeted control (such as directional gun adjustment), thereby suppressing splashing and drying in advance.

[0036] (3) Constructing a reliable technical closed loop that deeply integrates mechanism and data makes high-fidelity simulation the core of digital twin verification for online systems. Continuous data benchmarking ensures the universality and robustness of the method, solving the long-standing problem of disconnect between research and production. Compared with simply relying on industrial field measured data, introducing simulation data can generate theoretical samples covering various working conditions at low cost, making up for the lack of scarce event data; the data quality is controllable and the labels are accurate, providing internal flow field information that is difficult to obtain through actual measurement, enhancing the physical interpretability and predictive reliability of the model.

[0037] (4) The introduction of chaotic polynomial expansion (PCE) effectively addresses the random uncertainty of flame images and significantly improves the prediction robustness under smoke and dust interference. Moreover, the PCE method has the advantages of flexible parameter selection and small training amount. Even without simulation data, it can still achieve good prediction accuracy by training on only a small sample of data. It can be easily applied to the identification and prediction of flames at the mouth of other converters through a self-learning mechanism.

[0038] (5) It has low implementation costs, is easy to deploy in existing plants, and has strong scalability potential. Attached Figure Description

[0039] Figure 1 This is a simulation model diagram.

[0040] Figure 2 This is a velocity streamline contour map.

[0041] Figure 3 Example image of a flame with smoke and dust interference.

[0042] Figure 4 The image shows the edge distribution characteristics of the average optical flow velocity, illustrating that the average optical flow velocity follows a normal distribution.

[0043] Figure 5 The figure shows the marginal distribution characteristics of the standard deviation of brightness, illustrating that the standard deviation of brightness follows a Beta distribution.

[0044] Figure 6 This is a statistical comparison chart of the chaotic polynomial method and the Monte Carlo method, showing the comparison results of the two methods in terms of statistical indicators such as mean, standard deviation, and quantiles.

[0045] Figure 7 The figure shows a comparison of the cumulative distribution functions of the chaotic polynomial and the Monte Carlo method. The cumulative distribution function curves of the two methods are shown, demonstrating their basic overlap over the entire splash probability interval.

[0046] Figure 8Quantile-quantile plots for PCE and Monte Carlo methods are shown, illustrating a scatter plot comparing the quantiles of the two methods and demonstrating a high degree of consistency across the entire probability scale.

[0047] Figure 9 The feature importance analysis diagram based on the Sobol total effect index illustrates the contribution of each input feature to the uncertainty of splash probability prediction.

[0048] Figure 10 This is a flowchart of the steps of the present invention. Detailed Implementation

[0049] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Note: Unless otherwise specified, the experimental methods in the following examples are conventional methods, performed according to the techniques or conditions described in the literature in this field or according to the product instructions. Unless otherwise specified, the materials and reagents used in the following examples are commercially available.

[0051] This invention is the first to use the PCE method to process small sample data and perform uncertainty modeling. It can automatically analyze the probability distribution characteristics of input features and use orthogonal polynomial basis functions to perform proxy modeling of the mapping relationship between features and prediction targets, thereby achieving reliable prediction of splash probability while ensuring computational efficiency.

[0052] To address the dual challenges of difficult data acquisition and high uncertainty in chaotic polynomial modeling, a simulation model coupling fluid dynamics and radiative transfer is constructed. This model can simulate gas escape and flame evolution processes under different slag permeability states at the mechanistic level, thereby generating theoretical data covering multiple operating conditions. Simulation data not only effectively supplements and expands measured data when samples are insufficient, but also provides internal flow field information that is difficult to obtain directly in experiments, offering more physically meaningful input features for chaotic polynomial expansion modeling. The integration of simulation and experimentation forms a closed-loop technology driven by both "mechanism and data," improving the model's generalization ability and enhancing the interpretability and robustness of prediction results. Simulation data not only effectively supplements and expands measured data when samples are insufficient, but also provides internal flow field information that is difficult to obtain directly in actual measurements, offering more physically meaningful input features for chaotic polynomial expansion modeling. The integration of simulation and actual measurements forms a closed-loop technology driven by both mechanism and data, improving the model's generalization ability and enhancing the interpretability and robustness of prediction results. The flame features generated by simulation and those measured in the industrial field show a high degree of consistency in key indicators, with small statistical differences. This indicates that the constructed simulation model can accurately reproduce the flame evolution process under real molten slag conditions, and the deviation between simulation data and measured data is within an acceptable range. Therefore, simulation data is feasible as a substitute for measured data for model training and validation, providing a reliable guarantee for building high-precision prediction models in scenarios where measured samples are scarce.

[0053] The concept of this invention originates from a research path that involves deriving the mechanism through simulation, verifying the simulation with data, and utilizing chaotic polynomials to handle uncertainty. During the research and development process, the following technical challenges were addressed:

[0054] To address the random noise and uncertainty in flame images, a chaotic polynomial expansion (PCE) method is introduced to quantify and model the uncertainty of extracted features, effectively reducing the impact of smoke and dust interference on prediction accuracy.

[0055] like Figure 10 As shown, a method for analyzing converter slag condition and predicting splashing includes: Step 1. Construct a multiphysics simulation geometric model.

[0056] Establish as Figure 1 The converter geometry model shown includes the furnace body, molten pool surface, and furnace window outlet. The upper surface of the molten pool is discretized into multiple independent control regions, and a parameterized dynamic gas mass flow rate inlet boundary condition is set for each region i, with the functional form being: .in, The steady-state fundamental gas generation rate, permeability amplitude coefficient This is the primary distinguishing feature from existing technologies; it directly quantifies and controls the permeability state of the slag in that region. This is achieved by assigning different... This value allows for the construction of various spatially non-uniform slag conditions in simulations. This is a custom pulse width modulation function. It is a smooth start function. For example... Figure 1 As shown.

[0057] Step 2. Run the simulation and extract the quantitative indicators of flame characteristics.

[0058] After setting the initial conditions, material properties, and solver parameters, run the coupled simulation. Once the calculation converges, extract the simulation results for the furnace window region, including: velocity streamline contour maps (reflecting the airflow structure), such as... Figure 2 As shown. The mean curvature is extracted from the velocity streamlines. As a core quantitative indicator.

[0059] Step 3. Acquire and process industrial site images.

[0060] A high-temperature resistant industrial camera is installed on the side of the furnace window of the corresponding converter to continuously capture flame images throughout the entire blowing process. Flame images at key time points are extracted to create a real-world image library. For example... Figure 3 As shown.

[0061] Step 4. Feature processing and uncertainty modeling based on chaotic polynomial expansion.

[0062] This technology aims to address the difficulty of handling random uncertainties in flame images in existing technologies. It includes:

[0063] (4.1) Feature extraction: Multi-dimensional feature extraction is performed on the acquired flame images: Texture feature extraction: Based on the gray-level co-occurrence matrix (GLCM), entropy, energy, contrast, and inverse difference moment (Homogeneity) are extracted.

[0064] Color feature extraction: Extract the first moment (mean), second moment (variance), and third moment (skewness) of H, S, and V in the HSV space as statistics.

[0065] Dynamic feature extraction: Optical flow vectors between consecutive frames are extracted based on optical flow method, and the average optical flow velocity is statistically analyzed as dynamic feature.

[0066] Key physical feature extraction: Calculating the average curvature of the "pseudo-streamlines" derived from real-world flame images. Compare with simulation results.

[0067] (4.2) Feature selection and dimensionality reduction: The Random Forest Recursive Elimination (RF-RFE) method is used to select features and determine key input variables.

[0068] (4.3) Constructing a chaotic polynomial proxy model: Chaotic polynomial expansion is an efficient method for quantifying uncertainty. Assume the system performance response function is... The input features are random variables. The chaotic polynomial model will Expanded to:

[0069] in, It is an orthogonal polynomial basis. For expansion coefficients, This is the truncation order.

[0070] Marginal distribution analysis is performed on the selected features to determine the probability distribution type of each feature (normal distribution corresponds to Hemite orthogonal polynomial, Beta distribution corresponds to Jacobi orthogonal polynomial), and the coefficients are solved by least squares regression using the training dataset.

[0071] The training set is composed of the simulation data from step 1 and the measured labeled data from step 3, and is used as the training data for the chaotic polynomial surrogate model to complete the training of the chaotic polynomial surrogate model.

[0072] Step 5. Establish a permeability state identification model and splash prediction.

[0073] Through steps 1-4, multiple combinations of permeability parameters are obtained. →Simulation features → Actual characteristics A mapping database.

[0074] Calibrate the simulation model: Ensure and They are statistically highly correlated.

[0075] State recognition: Establishing a system based on real-world flame images for calculation By combining the probability distribution output by the chaotic polynomial model, the permeability parameters of each region of the current molten pool are inferred in reverse. (i.e., the identification model of slag state).

[0076] Using a chaotic polynomial surrogate model as a splash / drying predictor, in converter steelmaking, when the obtained permeability parameters are consistently low and fluctuate drastically, or when the splash / drying probability output by the chaotic polynomial surrogate model exceeds the threshold, the system issues a warning that the slag permeability is deteriorating and there is a risk of splash / drying.

[0077] To determine the accuracy of the prediction, when the system's output of splash / drying risk warnings does not match the actual operating conditions, a feedback mechanism is established to augment the model with data. The augmented data is used to train and optimize the chaotic multinomial surrogate model, improving the model's accuracy. If a splash / drying event actually occurs, but the system fails to issue a warning before the event occurs or the warning probability is lower than a set threshold, then the segment of flame image data is marked as a missed sample and added to the training set.

[0078] A converter slag state correlation analysis system includes a processor and a memory storing a computer program. When the processor executes the computer program, it implements the steps of the above-mentioned method for analyzing the state of converter slag and predicting splashing.

[0079] Example 1: Basic Process of Multiphysics Simulation and Fusion of Chaotic Polynomials Geometric Model Construction: Based on the actual dimensions of the target converter, a two-dimensional or three-dimensional geometric model including the furnace body, molten pool surface, and furnace window exit is established using multiphysics simulation software. The established multiphysics coupled mathematical model includes: the hydrodynamic field (NS equations) and the radiation transfer field.

[0080] Parameterized slag permeability: The upper surface of the molten pool is discretized into N independent sector regions, and a dynamic gas mass flow function is defined for each inlet. , .

[0081] Simulation calculation: Run the simulation and extract the average streamline curvature of the furnace window region. .

[0082] Image acquisition and feature extraction: Installing industrial cameras to acquire images such as... Figure 3 The flame image is shown. GLCM texture features, HSV color moments, optical flow velocity, and curvature are extracted. .

[0083] PEC Modeling: At a converter production site, an 8.9-megapixel industrial camera was used to collect images of the converter furnace mouth flame at a frequency of 26 frames per second. Image data of multiple splashing events were collected, and the image sequence within 5 seconds before the splashing occurred was extracted. Feature extraction was performed on the collected images: (1) Texture feature extraction: The gray-level co-occurrence matrix of each frame image was calculated, and four texture features were extracted: entropy, energy, contrast, and inverse difference matrix. (2) Color feature extraction: The image was converted from RGB space to HSV space, and the first, second, and third moments of the H, S, and V channels were extracted respectively, for a total of 9 color features. (3) Dynamic feature extraction: The optical flow vector between consecutive frames was calculated based on the optical flow method, and the average optical flow velocity of each frame image was statistically analyzed. The random forest recursive elimination method was used for feature screening. The key features after screening included: curvature, contrast, brightness standard deviation, entropy, saturation standard deviation, hue standard deviation, average hue, average optical flow velocity, and average saturation. Edge distribution fitting analysis was performed on the screened features, and the results are as follows. Figure 4 and Figure 5 As shown: the average optical flow velocity follows a normal distribution, while most other features (such as the standard deviation of brightness) follow a Beta distribution. For normally distributed features, Hermite orthogonal polynomials are chosen as the basis; for Beta distributed features, Jacobi orthogonal polynomials are chosen as the basis. A mixture of orthogonal polynomial basis functions is constructed, and the expansion coefficients are solved using the least squares method with the training dataset. The model is evaluated using 5-fold cross-validation: average coefficient of determination. Root mean square error This indicates that the model has strong nonlinear fitting ability and prediction accuracy. Based on the constructed chaotic multinomial surrogate model, the Monte Carlo simulation (PCE-MC) method is used to quantify the impact of uncertainty. Statistical results are as follows: Figure 6 As shown: PCE prediction mean: 0.7522, Monte Carlo mean: 0.7496, relative error: 0.35%; PCE prediction standard deviation: 0.5474, Monte Carlo standard deviation: 0.5408, relative error: 1.22%, Q95 quantile relative error: 1.04%, Q5 quantile relative error: 4.18%. The relative differences of all statistical indicators are less than 5%, verifying the reliability of the PCE method. The cumulative distribution function is compared as follows... Figure 7 As shown, the curves of the PCE and Monte Carlo methods largely overlap across the entire splash probability range, with only slight differences in the low-probability region, while they closely match in the high-probability region. This indicates that the PCE surrogate model possesses good statistical consistency and reliability. Quantile-quantile plots are shown below. Figure 8As shown, the scatter points are closely distributed along the diagonal, indicating that PCE and Monte Carlo methods have highly consistent statistical properties across the entire probability scale. Within the core probability interval, the scatter points almost entirely fall on the diagonal, demonstrating the accuracy of PCE in the main distribution region. The importance of each feature is analyzed using the Sobol index, and the results are as follows: Figure 9 As shown, the mean hue is the absolute dominant factor affecting the splash probability, with its Sobol index being much higher than other features. This indicates that color is the most critical physical characteristic determining splash behavior, and the model's prediction uncertainty mainly stems from hue. This aligns with the theory that an increase in CO concentration before splashing leads to changes in flame color. The trained PCE proxy model is deployed to the production site to extract features and predict splash probabilities from real-time acquired flame images, providing technical support for safe production in the converter steelmaking process.

[0084] Verification and Prediction: Comparison and Distribution. The splash probability is calculated using the PCE model, refer to... The status will be alerted.

[0085] Feedback Mechanism and Data Augmentation: When a splash / drying warning is issued, and the actual occurrence of splash / drying is not due to adjustments in the gun position, the relevant feature parameters of the image are labeled with the warning time and included in the training data. The model parameters are adjusted based on the augmented dataset, thus achieving self-learning. By monitoring abrupt changes in flame image features in real time, the system can accurately identify the occurrence of splash events. When splashing occurs, the system reviews the accuracy of its predictions prior to the splash. If the system's prediction was incorrect, the data prior to the splash is used as missed data to supplement and improve the training data, thereby enhancing the data's feature content.

[0086] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above content. Therefore, the scope of protection of the present invention should be defined by the appended claims.

Claims

1. A method for analyzing the state of converter slag and predicting splashing, characterized in that, include: Step 1. Construct a virtual-real combined image dataset using simulated geometric models and actual image acquisition and annotation; Step 2. Feature processing and uncertainty modeling based on chaotic polynomial expansion: Multi-dimensional feature extraction is performed on the flame images captured in step 1, and the random forest recursive elimination method is used for feature selection to determine the input variables. Construct a chaotic multinomial surrogate model, perform marginal distribution analysis on the selected features, determine the probability distribution type that each feature follows, and solve the chaotic multinomial coefficients using the training dataset through least squares regression. The simulation data and the measured labeled data from step 1 are used together to form a training set, which is used as the training data for the chaotic multinomial surrogate model, thus completing the training of the chaotic multinomial surrogate model. Step 3. Establish a permeability state identification model and splash / drying prediction: Obtain multiple sets of permeability parameter combinations →Simulation features → Actual characteristics The mapping relationship data is used to calibrate the simulation model; Establish a calculation based on real-shot flame images By combining the probability distribution output by the chaotic polynomial model, the combination of permeability parameters for each region of the current molten pool is inferred in reverse. The recognition model; Using a chaotic polynomial surrogate model as a splash / drying predictor, in converter steelmaking, when the obtained permeability parameters are consistently low and fluctuate drastically, or when the splash / drying probability output by the chaotic polynomial surrogate model exceeds the threshold, the system issues a warning that the slag permeability is deteriorating and there is a risk of splash / drying.

2. The method according to claim 1, characterized in that, The simulation geometric model described in step 1 is established through a multiphysics field model that couples dynamics and radiation transfer equations.

3. The method according to claim 1, characterized in that, In step 1, the camera used for actual image acquisition has more than 8 megapixels, and the acquisition frequency is no less than 20 frames per second.

4. The method according to claim 1, characterized in that, Step 2, feature extraction, includes texture features, color features, and dynamic features. Texture feature extraction involves extracting entropy, energy, contrast, and inverse difference matrix based on the gray-level co-occurrence matrix. Color feature extraction involves extracting the first, second, and third moments of H, S, and V in the HSV space. Dynamic feature extraction involves extracting the optical flow vector between consecutive frames based on the optical flow method and calculating the average optical flow velocity.

5. The method according to claim 1, characterized in that, In step 2, the chaotic polynomial expansion transforms the system performance response function. Expand as ,in, It is an orthogonal polynomial basis. For expansion coefficients, To truncate the order, orthogonal polynomials are used to expand the random features in order to quantify the uncertainty.

6. The method according to claim 1, characterized in that, Step 2 involves using the Sobol index to analyze feature importance and screen out the key features that contribute the most to splash / drying prediction.

7. The method according to claim 1, characterized in that, This also includes establishing feedback and data augmentation mechanisms: To determine whether the prediction is accurate, when the system's output of splash / re-drying risk warning does not match the actual working conditions, a feedback mechanism is established to augment the model with data. Augmented data is used to train and optimize the chaotic multinomial surrogate model, thereby improving the model's accuracy.

8. The method according to claim 7, characterized in that, If a splashing / drying event actually occurs, but the system fails to issue a warning before the event occurs or the warning probability is lower than the set threshold, then the flame image data segment is marked as a missed sample and added to the training set.

9. A converter slag state correlation analysis system, comprising a processor and a memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method for analyzing converter slag condition and predicting splashing as described in any one of claims 1-8.

Images