Converter end point carbon content and temperature online prediction method based on multi-modal data

Abstract

The invention relates to the technical field of ferrous metallurgy, and discloses a converter endpoint carbon content and temperature online prediction method based on multi-modal data. The method comprises the following steps: collecting visual image data, flame spectrum data and blowing process parameter data of furnace mouth flames in the blowing process of a converter to form multi-modal data; accurately positioning and segmenting a flame region by using a target detection and segmentation model, and extracting visual features and spectral features corresponding to the space according to the flame region; the visual features, the spectral features and the process parameters are fused and preprocessed and then input into a pre-trained symbol regression prediction model, and an end point carbon content prediction value and an end point temperature prediction value are obtained; on-line operation guidance information for adjusting the process parameters is generated based on the deviation between the predicted value and the target value and the input parameter sensitivity. According to the method, the end point prediction precision and interpretability are improved, the online control requirement of the converter is met, the end point double hit rate of the converter is effectively improved, and a closed loop from intelligent prediction to intelligent control is achieved.

Description

Technical Field

[0001] This application relates to the field of iron and steel metallurgy technology, and in particular to an online prediction method for converter endpoint carbon content and temperature based on multimodal data. Background Technology

[0002] In converter steelmaking, the final carbon content and molten steel temperature are key indicators for evaluating the control level of the blowing endpoint, directly affecting subsequent refining load, alloy yield, finished product quality, and production rhythm. Due to the characteristics of converter blowing processes such as strong multiphase reactions, intense heat and mass transfer, and large fluctuations in operating conditions, the final carbon content and temperature change rapidly over time and are influenced by multiple factors such as raw material conditions, oxygen supply system, bottom blowing conditions, and operating methods. Therefore, on-site online prediction and control of the endpoint state are usually required during the blowing process.

[0003] Existing methods for converter endpoint control and prediction mainly include: mechanistic modeling methods based on material and energy balance, statistical / machine learning prediction methods based on historical data, and visual or spectral detection methods based on furnace mouth flame characterization information. Mechanistic modeling methods rely on relatively complete process mechanisms and parameter identification. When faced with raw material fluctuations, changes in reaction paths, and on-site disturbances, model parameters are difficult to correct accurately in real time, affecting prediction accuracy and applicability. Data-driven prediction methods (such as neural networks) can obtain certain predictive capabilities using historical data, but the models often lack interpretability, making it difficult to clarify the influence relationship of various process parameters on the prediction results. On-site operators find it difficult to translate the prediction results into actionable process adjustment strategies. Detection methods based on flame images or flame spectra can reflect the blowing stage and reaction intensity, but in actual industrial environments, the furnace mouth is subject to smoke, background radiation, furnace lining interference, and drastic changes in flame morphology. The flame area is difficult to locate and segment stably and accurately, causing the extracted visual or spectral features to be easily mixed with non-flame information, thus affecting prediction stability and accuracy. At the same time, single-modal information is usually insufficient to fully characterize the converter endpoint state and cannot simultaneously meet the requirements for predicting carbon content and temperature. In addition, some existing methods still have shortcomings in terms of real-time performance, robustness, and guidance for on-site operations, making it difficult to meet the comprehensive requirements of "accuracy, stability, interpretability, and guidance" for converter online control.

[0004] Therefore, how to construct a converter endpoint intelligent prediction and control method that can deeply integrate multi-source sensing information, ensure high accuracy and strong interpretability of model prediction, and directly drive online process optimization, thereby improving the accuracy, stability and operability of endpoint control, has become a key technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides an online prediction method for converter endpoint carbon content and temperature based on multimodal data, aiming to solve technical problems such as insufficient accuracy of traditional mechanistic models, poor interpretability of data-driven models, insufficient fusion of multi-source information, and lack of online decision guidance capabilities, thereby achieving an upgrade from intelligent sensing to closed-loop control.

[0006] In a first aspect, this application provides an online prediction method for converter endpoint carbon content and temperature based on multimodal data, the method comprising: S1. Collect visual image data of the furnace mouth flame, spectral data of the furnace mouth flame, and process parameter data of the converter blowing process to form multimodal data. S2. Input the visual image data into the target detection and segmentation model to locate and segment the flame area at the furnace opening, and obtain the spatial location information of the flame area. S3. Based on spatial location information, extract visual features from the segmented flame region and extract spectral features corresponding to the space of the flame region from spectral data; S4. The visual features, spectral features and process parameter data are fused and preprocessed to construct a fused feature vector; S5. Obtain the pre-trained symbolic regression prediction model, input the fused feature vector into the symbolic regression prediction model, and obtain the predicted values ​​of the endpoint carbon content and the endpoint temperature. S6. Based on the deviation between the predicted value and the target value, and the sensitivity of each input parameter under the current operating conditions, generate online operation guidance information for adjusting process parameters; S7. Collect actual endpoint test values, update the coefficients of the symbolic regression prediction model online in a sliding window manner, and switch to conservative control mode when the confidence level is lower than the preset confidence threshold.

[0007] In conjunction with the first aspect, in the first implementation of the first aspect of this application, the process parameter data includes at least the weight of molten iron, the temperature of molten iron, the chemical composition of molten iron, the amount of scrap steel added, the amount of flux added, the total amount of oxygen blown, the oxygen blowing intensity, the oxygen blowing time, the oxygen lance height, the bottom blowing gas flow rate, the current blowing time, and the furnace age.

[0008] In conjunction with the first aspect, in the second implementation of the first aspect of this application, the target detection and segmentation model is a flame segmentation model based on deep learning, which outputs a pixel-level segmentation mask and / or bounding box coordinates of the flame region. The target detection and segmentation model is trained using a flame image dataset covering the early, middle, severe decarburization and final stages of converter blowing.

[0009] In conjunction with the first aspect, in the third implementation of the first aspect of this application, the visual features include at least three of the following: color features, texture features, morphological features, and dynamic change features of the flame region.

[0010] In conjunction with the first aspect, in the fourth implementation of the first aspect of this application, the color features include statistical features of the H, S, and V components extracted from the HSV color space, the texture features include statistical texture features extracted based on the gray-level co-occurrence matrix, the morphological features include the area, aspect ratio, and roundness of the flame region, and the dynamic change features include flame brightness fluctuation features and / or flame region area fluctuation features.

[0011] In conjunction with the first aspect, in the fifth implementation of the first aspect of this application, the spectral features include characteristic element spectral line intensity features, continuous spectrum features, and spectral line ratio features. The characteristic element spectral line intensity features include the integrated intensity of spectral lines of at least two elements selected from potassium, sodium, calcium, and iron. The continuous spectrum features include the spectral centroid and spectral width. The spectral line ratio features include the potassium-sodium intensity ratio and / or the calcium-iron intensity ratio.

[0012] In conjunction with the first aspect, in the sixth implementation of the first aspect of this application, the fusion and preprocessing in S4 includes standardizing and / or normalizing the fused features, and using a dimensionality reduction method to reduce the dimensionality of the fused feature vector to reduce feature correlation.

[0013] In conjunction with the first aspect, in the seventh implementation of the first aspect of this application, the symbolic regression prediction model is trained by the symbolic regression method of genetic programming. During the training process, a multi-objective optimization strategy is adopted to minimize the prediction error and the complexity of the expression, and a penalty term is set for the complexity of the expression to obtain a prediction model with a simple structure.

[0014] In conjunction with the first aspect, in the eighth implementation of the first aspect of this application, metallurgical process constraints are introduced when training the symbolic regression prediction model. The metallurgical process constraints include at least directional constraints and boundary constraints. The directional constraints are used to ensure that the direction of influence of variables related to oxygen blowing on the endpoint carbon content conforms to the laws of metallurgical physics, and the boundary constraints are used to limit the predicted values ​​of the endpoint carbon content and / or the predicted values ​​of the endpoint temperature within a preset physical reasonable range.

[0015] In conjunction with the first aspect, in the ninth implementation of the first aspect of this application, the sensitivity is obtained by taking the partial derivative of the symbolic regression prediction model with respect to each input parameter, and the partial derivative result is standardized. The online operation guidance information determines the direction and amount of adjustment for oxygen blowing volume, coolant addition volume, and / or oxygen lance height by combining the deviations of the predicted values ​​of the endpoint carbon content and the endpoint temperature from the target value, under the condition of meeting the constraints of process parameter adjustment range.

[0016] Compared with the prior art, the beneficial effects of the technical solution of this application are at least as follows: 1. Significantly Improved Prediction Accuracy: By integrating three types of heterogeneous data—visual, spectral, and process parameter data—and utilizing advanced visual models such as YOLOv8, precise segmentation of the flame region was achieved, ensuring the extraction of "pure" and effective features from the original signal. This deep fusion and synergistic utilization of multimodal information overcomes the shortcomings of incomplete information from a single data source, laying a solid foundation for high-precision modeling and improving the accuracy and stability of online prediction of endpoint carbon content and temperature.

[0017] 2. Excellent model interpretability and strong guidance: Abandoning traditional "black box" neural networks, the model employs symbolic regression to directly generate readable mathematical expressions. These expressions clearly reveal the quantitative relationship between input parameters and the prediction target. Based on this, the sensitivity (partial derivatives) of each process parameter can be easily calculated, thus transforming the model's prediction results into parameter weight analysis that directly guides on-site operations, greatly enhancing engineers' trust and control over the model's decision-making process.

[0018] 3. Achieves an intelligent closed loop from prediction to optimized control: This invention goes beyond simply predicting the endpoint state; it further constructs an online decision-making model based on multi-objective optimization. This model comprehensively considers the deviation between the carbon and temperature objectives, as well as the costs and constraints of adjusting various parameters, automatically solving for the optimal process parameter adjustment scheme (such as oxygenation rate). This achieves a leap from "post-event prediction" to "in-event control," forming a complete intelligent closed loop of "perception-prediction-interpretation-decision-execution," directly serving production operations and effectively improving the endpoint "double hit rate."

[0019] 4. Strong robustness and adaptability: Prior knowledge of metallurgical processes, such as directional constraints, is incorporated into the model construction to ensure that the symbolic regression results conform to physical laws, thereby improving the model's generalization ability and physical rationality. Simultaneously, the system supports online feedback and periodic updates based on actual smelting results, enabling it to adapt to operating conditions such as raw material fluctuations and equipment status changes, and possessing adaptive optimization capabilities for long-term service.

[0020] 5. Low implementation cost: It can be implemented using conventional equipment such as industrial cameras and spectrometers, and the effectiveness of feature extraction is improved by obtaining the flame position and coordinates through precise segmentation, achieving high-precision prediction and guidance without significantly increasing hardware costs. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the online prediction method for converter endpoint carbon content and temperature based on multimodal data according to this application; Figure 2 This is a schematic diagram of the system architecture and data flow of the method in this application. Detailed Implementation

[0023] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms “comprising” or “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0024] For ease of understanding, the specific process of the embodiments of this application is described below. Figure 1 The diagram shows a flowchart of the online prediction method for converter endpoint carbon content and temperature based on multimodal data provided by the present invention. The flowchart specifically includes the following steps: S1. Collect visual image data of the furnace mouth flame, spectral data of the furnace mouth flame, and process parameter data of the converter blowing process to form multimodal data.

[0025] In one specific embodiment, the process parameter data includes at least the weight of molten iron, the temperature of molten iron, the chemical composition of molten iron, the amount of scrap steel added, the amount of flux added, the total amount of oxygen blown, the oxygen blowing intensity, the oxygen blowing time, the oxygen lance height, the bottom blowing gas flow rate, the current blowing time, and the furnace age.

[0026] Specifically, an industrial camera and spectrometer are installed at the furnace opening, and a unified time reference is established so that image frames and spectral curves are written into data records under the same furnace number and acquisition time index. Images are continuously acquired at a predetermined resolution and frame rate, and spectra are continuously acquired at a set band and resolution, with dark current subtraction, wavelength calibration, and noise smoothing completed. Then, each spectral sample is mapped to an adjacent image frame by timestamp alignment, forming a one-to-one correspondence between "image data and spectral data". This avoids the problem of non-flame signals caused by furnace opening brick lining, smoke, and background radiation being mistakenly taken as valid input.

[0027] The process parameters are synchronously read from the PLC / MES and organized according to "static quantity per furnace + time sequence quantity". The process parameter data includes molten iron composition, scrap steel addition amount, oxygen blowing amount, oxygen blowing time, flux addition amount, etc. Specifically, it includes: raw material parameters, blowing process parameters, and process parameters. Among them, raw material parameters include molten iron parameters (such as molten iron weight (tons), molten iron temperature (°C), etc.), molten iron chemical composition (such as carbon content (%), silicon content (%), manganese content (%), phosphorus content (%), sulfur content (%), etc.), scrap steel parameters (such as total scrap steel weight (tons), etc.), scrap steel type ratio (such as heavy scrap steel ratio (%), light scrap steel ratio (%), pig iron block ratio (%), etc.), flux parameters (such as lime addition amount (kg), light calcined dolomite addition amount (kg), fluorite addition amount (kg), etc.). Ore addition amount (e.g., sinter (kg), pellets (kg), lump ore (kg), etc.); blowing process parameters include oxygen supply parameters (e.g., total oxygen blowing amount (Nm³), oxygen blowing time (min), oxygen supply intensity (Nm³ / min), oxygen lance height mode (high-low-high or constant lance position), etc.) and bottom blowing parameters (e.g., equal bottom blowing gas type (Ar, N2), bottom blowing gas flow rate (Nm³ / h)); process parameters include time parameters (e.g., start-up time, current blowing time, expected end time, etc.) and equipment parameters (e.g., converter age (heating), furnace lining condition assessment value, etc.).

[0028] The parameter tables after unit conversion, missing value imputation, and outlier removal are stitched together into the corresponding image-spectral record according to the same index to obtain a multimodal sample set. This provides a common input for subsequent extraction of flame features based on spatial location information and fusion with process parameters, thereby alleviating the problem of insufficient information and unstable prediction caused by relying on only a single mode.

[0029] S2. Input the visual image data into the target detection and segmentation model to locate and segment the flame area at the furnace opening, and obtain the spatial location information of the flame area.

[0030] In one specific embodiment, the target detection and segmentation model is a deep learning-based flame segmentation model that outputs pixel-level segmentation masks and / or bounding box coordinates of the flame region. The target detection and segmentation model is trained using a dataset of flame images covering the early, middle, severe decarburization, and final stages of converter blowing.

[0031] Specifically, the acquired visual images are processed based on the YOLOv8 segmentation model to identify and segment the flame region in the flame image. The visual images are entered into the image buffer according to the furnace number and timestamp and the resolution is resampled. Each frame of the image is used as the model input with a fixed size, and the bounding box coordinates and pixel-level segmentation mask of the flame region are output as precise location information.

[0032] The YOLOv8 segmentation model is trained using a dedicated converter flame dataset, which includes flame images from different smelting stages. During the training phase, sample pairs are established between each frame image in the labeled dataset and its pixel-level segmentation annotations. The dataset covers flame morphology in the early, middle, severe decarburization, and final stages of smelting. Enhancement methods such as Mosaic, rotation, brightness and contrast jitter, and Gaussian noise are combined to expand the sample distribution, thereby addressing the flame boundary instability caused by furnace lining bricks, smoke, and background radiation, and reducing the regional drift wind generated by rule-based methods such as threshold segmentation under changing operating conditions. During the inference phase, each frame of image is input into the flame segmentation model, which outputs the segmentation mask matrix and bounding box coordinate vector corresponding to the same frame. The segmentation mask marks the set of pixels in the flame region in the pixel coordinate system, and the bounding box gives the region range in the form of the upper left corner coordinates and width and height. The two, together with the timestamp of the frame, are written into the spatial location information record, thus forming a correspondence of "image frame index - mask - bounding box" and providing regional constraints for subsequent feature extraction. This ensures that visual features are only statistically analyzed within the mask coverage area, and spectral features can be extracted spatially based on the bounding box coordinates. This avoids signals from non-flame regions from entering the feature set and causing prediction bias. At the same time, the model processing time meets the online timeliness requirements. Preferably, the specific training configuration for the YOLOv8 segmentation model is as follows: Input image resolution: Adjusted to 640x640 pixels to meet real-time processing speed requirements; Training cycle: 300 epochs, automatically stopped when the accuracy of the validation set no longer improves for 10 consecutive epochs; Batch size: 16, dynamically adjusted based on GPU memory; Optimizer: AdamW is used with an initial learning rate of 0.01 and a cosine annealing scheduler is applied; Data augmentation: Mosaic augmentation, random rotation (±10°), brightness and contrast jitter (±20%), and Gaussian noise were employed to improve the robustness of the model in complex industrial environments. Dedicated dataset: Contains no fewer than 5,000 labeled images, covering flame morphology in the early, middle, intense decarburization, and final stages of smelting, ensuring that the model has the ability to accurately identify flames throughout the entire smelting cycle.

[0033] S3. Based on spatial location information, extract visual features from the segmented flame region and extract spectral features corresponding to the space of the flame region from the spectral data.

[0034] In one specific embodiment, the visual features include at least three of the following: color features, texture features, morphological features, and dynamic change features of the flame region.

[0035] In one specific embodiment, the color features include statistical features of the H, S, and V components extracted from the HSV color space; the texture features include statistical texture features extracted based on the gray-level co-occurrence matrix; the morphological features include the area, aspect ratio, and roundness of the flame region; and the dynamic change features include flame brightness fluctuation features and / or flame region area fluctuation features.

[0036] In one specific embodiment, the spectral features include characteristic element spectral line intensity features, continuous spectrum features, and spectral line ratio features. The characteristic element spectral line intensity features include the integrated intensity of spectral lines of at least two elements selected from potassium, sodium, calcium, and iron. The continuous spectrum features include the spectral centroid and spectral width. The spectral line ratio features include the potassium-sodium intensity ratio and / or the calcium-iron intensity ratio.

[0037] Specifically, the segmentation mask obtained in step S2 is bound to the bounding box coordinates and the image frame timestamp to form a spatial location information record. The set of pixels in the flame area is limited by the mask rather than by the entire frame, so that the pixels corresponding to the furnace background, smoke and dust obstruction and furnace lining reflection do not enter the subsequent feature calculation and reduce feature drift caused by the introduction of non-flame signals.

[0038] Within the same frame, the pixels of the flame region are converted to the HSV color space. The mean and variance of the H component (hue), the mean of the S component (saturation), and the mean, maximum, and 95th quantile of the V component (luminance) are calculated on the pixels covered by the mask to obtain color features. After graying the flame region, the mean values ​​of contrast, energy, correlation, and homogeneity are calculated on gray-level co-occurrence matrices at pixel distances of 1 and angles of 0°, 45°, 90°, and 135° to obtain texture features. The pixel area of ​​the flame region, the aspect ratio of the bounding rectangle, and the ratio of the perimeter to the area (circularity) of the outline are calculated on the mask contour to obtain morphological features. A five-frame window is formed at 10 frames per second on the time axis. The standard deviation of the mean V component within the window is calculated as the luminance fluctuation rate, and the moving average of the pixel area change rate is calculated as the area fluctuation rate to obtain dynamic change features. These visual features maintain a one-to-one correspondence with the corresponding frame indices.

[0039] The spectral side selects synchronously acquired spectral curves with the same timestamp as the frame and performs flame spatial correspondence verification based on spatial location information. This ensures that the spectral feature source and visual feature point to the same flame region, rather than spectral distortion caused by mixing in non-flame fields of view. The spectral lines of potassium (K 769nm), sodium (Na 589nm), calcium (Ca 422nm), and iron (Fe 371nm) are integrated within a window of ±0.2nm around the peak value on the selected spectral curve to obtain the spectral line intensities of at least two elements, thus resisting noise fluctuations. Further calculations are made of the spectral centroid, spectral width, and full width at half maximum (FWHM) to characterize the continuous spectrum. The K / Na intensity ratio and Ca / Fe intensity ratio are calculated as ratio features, and default values ​​are used when the denominator intensity is below a threshold to avoid division by zero and noise amplification. These spectral features maintain a correspondence with the same frame index and can jointly characterize the blowing stage and reaction intensity with visual features. This addresses the problems of unstable flame region extraction under strong interference environments and prediction fluctuations caused by difficulty in aligning multi-source information in existing technologies, ensuring that subsequent fusion inputs have consistent spatial sources and temporal correspondence.

[0040] S4. The visual features, spectral features and process parameter data are fused and preprocessed to construct a fused feature vector.

[0041] In one specific embodiment, the fusion and preprocessing in S4 includes standardizing and / or normalizing the fused features, and using a dimensionality reduction method to reduce the dimensionality of the fused feature vector to reduce feature correlation.

[0042] Specifically, visual feature records, spectral feature records, and process parameter records are aligned with the furnace number and timestamp to form the same sample index. Multidimensional statistics obtained from the visual side, spectral line intensities and continuous spectral quantities obtained from the spectral side, and oxygen supply and feeding values ​​obtained from the process side are concatenated under this index to form a joint feature sequence, maintaining the correspondence of "the same index corresponding to the same operating condition at the same time." After concatenation, standardization and / or normalization are performed on each dimension of the features to eliminate weight bias caused by different units and value ranges, and to suppress the influence of sensor amplitude drift on the model input. Subsequently, the feature covariance structure is calculated using the sample set, and principal component analysis is used for dimensionality reduction. Highly correlated feature combinations are projected into a low-dimensional space while retaining a preset variance ratio, ensuring that the dimensionality-reduced vectors still correspond one-to-one with the original index and serve as input to the symbolic regression model. This addresses the problems of strong feature correlation, unstable model training, and prediction fluctuations caused by insufficient fusion resulting from the direct superposition of multi-source information, enabling the input to express the relationship between images, spectra, and process parameters under conditions of consistent scale and reduced correlation.

[0043] S5. Obtain the pre-trained symbolic regression prediction model, input the fused feature vector into the symbolic regression prediction model, and obtain the predicted values ​​of the endpoint carbon content and the endpoint temperature.

[0044] In one specific embodiment, the symbolic regression prediction model is trained using a symbolic regression method based on genetic programming. During the training process, a multi-objective optimization strategy is employed to minimize both the prediction error and the complexity of the expression, and a penalty term is set for the complexity of the expression to obtain a prediction model with a concise structure.

[0045] In a preferred embodiment, metallurgical process constraints are introduced when training the symbolic regression prediction model. The metallurgical process constraints include at least directional constraints and boundary constraints. The directional constraints are used to ensure that the direction of influence of variables related to oxygen blowing on the endpoint carbon content conforms to the laws of metallurgical physics, and the boundary constraints are used to limit the predicted values ​​of the endpoint carbon content and / or the predicted values ​​of the endpoint temperature within a preset physical reasonable range.

[0046] Specifically, the fusion feature vector is correlated with the measured carbon content and measured temperature at the furnace endpoint by furnace number to form a training sample set. During the training phase, the symbolic regression prediction model searches for expression structures using genetic programming and combines and maps the fusion features using the function set {+, -, ×, ÷, √, log, exp, sin, cos, tanh, relu}. Individuals in the population iteratively evolve under crossover and mutation operations, and the normalized root mean square error is used as the fitting index. At the same time, a Pareto multi-objective mechanism is introduced to balance error and expression complexity, and a penalty coefficient is added to the complexity to suppress expression inflation. This ensures that the model has an executable computational load during online control and avoids excessive structural complexity caused by pursuing only a fit, making it difficult to use for on-site interpretation and parameter tuning.

[0047] For example, the population size is set to 2500–3500 individuals, ensuring sufficient genetic diversity; the number of generations is set to 120–180, ensuring sufficient convergence; the selection mechanism uses a size of 5–8; the crossover probability of subtrees for crossover and mutation is 0.8–0.9, and the point mutation probability is 0.08–0.12; the stopping criterion is: when the optimal fitness improvement is less than 1 × 10 for 20 consecutive generations. -6 Training automatically stops when the time is right; the initialization method is a "semi-growing method," with an initial tree depth of 4-7 to balance diversity and convergence speed; the objective function is to minimize the normalized root mean square error, incorporating Pareto optimization to balance model accuracy and complexity. The complexity penalty term is... Penalty coefficient The range of 0.001 to 0.01 was determined through grid search; the symbolic regression employed multi-objective optimization, taking into account both prediction accuracy and formula simplicity.

[0048] Metallurgical process constraints are introduced during training to address potential issues of inconsistent physical orientation and prediction out-of-bounds errors in data-driven models. Directional constraints are implemented by encapsulating and structurally restricting oxygen blowing-related inputs during the expression construction phase, ensuring their contribution to the carbon content expression remains negative and consistent with the decarburization mechanism. Preferably, directional constraints are implemented through a custom function encapsulation, defining... ,in, To prevent gradient explosion at the minimum value, 'x' represents the input feature value corresponding to the oxygen blowing rate parameter, which is the dimension representing the oxygen blowing rate in the fused feature vector. When constructing the expression, the oxygen blowing rate parameter is forced through this function to ensure it consistently contributes negatively to the carbon content expression. Boundary constraints post-process and trim the model output to limit the predicted carbon content and temperature values ​​from falling within a preset physical range. The fitness function can also incorporate a dimensional consistency penalty term to favor expressions with reasonable dimensions. Metallurgical process constraints can also include a dimensional consistency bias, performing a dimensional check on the expression during symbolic regression search, penalizing or reducing the weight of combinations with inconsistent dimensions to ensure the optimal expression is compatible in terms of input and output physical dimensions. While strict dimensional consistency is not mandatory, expressions with reasonable dimensions are preferred.

[0049] The symbolic regression prediction model includes a trained sub-model for predicting carbon content and a sub-model for predicting temperature, which are used to output the predicted values ​​of the final carbon content and the final temperature, respectively. The model is constructed by training the carbon content prediction model and the temperature prediction model separately, ensuring that the same fused feature vector yields two types of outputs on two expression paths while maintaining the correspondence between "the same operating condition at the same time and the two types of final states." In online application, the fused feature vector is input into the trained model, and the predicted values ​​of the final carbon content and the final temperature are directly calculated. Furthermore, the expression structure supports explicit expression of the influence of input variables to support subsequent sensitivity calculations based on partial derivatives and correlates prediction deviations with adjustable process parameters such as oxygen blowing rate, feed rate, and lance position. This addresses the control risks arising from the difficulty of providing executable parameter tuning basis in black-box models and the deviation of data-driven results from metallurgical principles.

[0050] For example, the carbon content prediction model expression is as follows: The temperature prediction model expression is: ,in, , For the specific mathematical expression obtained by symbolic regression, X1, X2, and X3 are the components of the fused feature vector after fusion and preprocessing. They usually correspond to the first three feature components after dimensionality reduction. X1 is the first dimension of the fused feature vector (often the first principal component), which is formed by a linear combination of visual features, spectral features, and process parameters according to weights. X2 is the second dimension of the fused feature vector (often the second principal component), which is also a combination of the above multimodal features. X3 is the third dimension of the fused feature vector (often the third principal component), which also comes from a combination of multimodal features.

[0051] When applied online, the current fused feature vector is input into the trained symbolic regression prediction model, that is, input into the carbon content prediction expression and the temperature prediction expression respectively, to obtain the corresponding endpoint carbon content prediction value and endpoint temperature prediction value.

[0052] Figure 2 This diagram illustrates the system architecture and data flow of the method described in this invention. As shown, the system mainly includes a multimodal data acquisition module, a flame region localization and segmentation module, a visual and spectral feature extraction module, a feature fusion module, a symbolic regression prediction module, and a result output module. An industrial camera and spectrometer simultaneously acquire images and raw spectral data of the furnace flame. The YOLOv8 model processes the images to achieve precise localization and segmentation of the flame region and outputs its coordinates (ROI). Based on these coordinates, visual features are extracted from the segmented image regions, and spectral features are extracted from the corresponding spectral data. The extracted visual and spectral features are fused with process parameters from the production system to form a fused feature vector. This vector is input into a pre-trained symbolic regression model, ultimately outputting the predicted carbon content and temperature at the converter endpoint. The arrows in the diagram indicate the main data flow directions.

[0053] S6. Based on the deviation between the predicted value and the target value, and the sensitivity of each input parameter under the current operating conditions, generate online operation guidance information for adjusting process parameters.

[0054] In one specific embodiment, the sensitivity is obtained by taking the partial derivatives of the symbolic regression prediction model with respect to each input parameter, and then standardizing the partial derivative results; The online operation guidance information determines the direction and amount of adjustment for oxygen blowing volume, coolant addition volume, and / or oxygen lance height by combining the deviations of the predicted values ​​of the endpoint carbon content and the endpoint temperature from the target value, under the condition of meeting the constraints of process parameter adjustment range.

[0055] Specifically, in online applications, after obtaining the corresponding predicted values ​​of the endpoint carbon content and the endpoint temperature, the deviations are calculated by subtracting them from the target carbon content and the target temperature, respectively. Furthermore, at the same operating point, the partial derivatives of the two symbolic regression expressions (the carbon content prediction expression and the temperature prediction expression) are calculated with respect to each input parameter to obtain the local sensitivity of each parameter to the carbon content output and the temperature output. The sensitivity is then standardized according to the standard deviation of the parameter and the output in historical samples to eliminate differences in dimensions and amplitudes, so that the influence of different parameters can be compared on the same scale.

[0056] Under the condition that deviation is the driving constraint and standardized sensitivity is the basis for direction and intensity, a solution process for adjustment is constructed with the common convergence of carbon deviation and temperature deviation as the objective. Adjustable process parameters are selected as oxygen blowing rate, coolant addition rate, and oxygen lance height, and single adjustment upper and lower limits and damping coefficient constraints are applied to avoid adjustment overshoot. For example, the single oxygen adjustment rate is limited to not exceeding a preset threshold, and a damping coefficient is set in the range of 0.1 to 0.5 to balance response speed and stability. The adjustment direction and adjustment amount of each parameter are solved and output as operation guidance information. When carbon is high and oxygen blowing rate sensitivity is negative, the oxygenation direction is matched with the deviation and the recommended oxygenation rate is calculated according to the ratio of deviation to sensitivity. At the same time, the temperature rise effect caused by oxygenation is compensated by combining temperature deviation and temperature sensitivity. When the temperature approaches the upper limit, the coolant addition rate or lance position adjustment is given in linkage to suppress the temperature rise. Thus, the prediction results are transformed into executable process adjustments and the control difficulties caused by only providing prediction values ​​without operational basis are reduced.

[0057] For example, the global importance weights of each parameter are calculated based on historical data. Based on the mathematical expression obtained from symbolic regression, the partial derivatives of each input parameter with respect to the prediction results are calculated to obtain the sensitivity of each parameter under the current operating conditions. The standardized calculation of the sensitivity is as follows: , and Parameters and output The standard deviation in historical data. The confidence quantification formula is: ,in, The root mean square error of the model on the most recent 50-heat sliding window validation set. This represents the normal range of variation for the target variable. A larger value indicates a more reliable prediction result.

[0058] Establish a prediction confidence assessment mechanism, triggering manual intervention when the confidence level falls below a threshold. For example, a tiered early warning mechanism (setting three warning levels) could be implemented: It provides a "Caution" prompt, suggesting that operators pay close attention; This triggers manual intervention, requiring operator confirmation. It automatically switches to conservative control mode, increasing the damping coefficient to 0.5.

[0059] The process of generating operational guidance information involves: establishing a multi-objective optimization decision-making model with process parameter adjustments as variables; and defining the adjustment vector. To guide the operation, solve the following constrained optimization problem: The constraints are , These are the standard errors of the carbon content prediction model and the temperature prediction model. It is the damping coefficient (with a value range of 0.1 to 0.5). S is the process parameter adjustment vector. C and S T Let ΔC be the sensitivity vector of carbon content and temperature to process parameters, respectively. target and ΔT target These are the target adjustments required for carbon content and temperature, respectively. These are the upper and lower limits of the adjustment amount (e.g., a maximum single oxygen adjustment of 500 Nm³). Solving this optimization problem yields the optimal process parameter adjustment amount, which serves as operational guidance information.

[0060] S7. Collect actual endpoint test values, update the coefficients of the symbolic regression prediction model online in a sliding window manner, and switch to conservative control mode when the confidence level is lower than the preset confidence threshold.

[0061] Specifically, after each heat of the converter is stopped, the endpoint test value and temperature value obtained from the endpoint sampling and temperature measurement are written into the heat database as a monitoring signal, and a one-to-one correspondence is established with the fusion feature vector formed at the corresponding time before the heat is stopped, so that each sample simultaneously contains a group record of "fusion feature vector - measured endpoint carbon content - measured endpoint temperature - predicted value of model output".

[0062] During online operation, the symbolic regression prediction model maintains the expression structure unchanged while setting the adjustable coefficients in the expression as parameters to be updated. The most recent preset number of furnaces forms a sliding window training set, and the coefficients are incrementally updated with the measured values ​​within the window as the target of the prediction error. The coefficient update can be achieved by solving the least squares with regularization terms or recursive least squares on the samples within the window, so that the newly entered furnaces contribute to the update of the coefficients and the earliest furnaces no longer participate in the update after leaving the window. This adapts to the data distribution drift caused by changes in furnace lining status, raw material fluctuations and oxygen supply conditions. At the same time, the update process uses post-processing with boundary constraints to avoid the output falling into non-physical ranges. The sliding window length can be taken from the most recent 50 heats, and the root mean square error on the validation set of this window is used as the online health measure of the model. The prediction confidence is calculated in combination with the normal range of change of the target variable. The confidence decreases as the window error increases and is used to trigger graded early warning. When the confidence is lower than the preset confidence threshold, the control strategy is switched to conservative control mode. The damping coefficient is increased to suppress the adjustment of process parameters, so that the online operation guidance changes from "automatic execution" to "manual intervention / conservative execution" in low confidence scenarios. This addresses the problems of model mismatch caused by non-stationarity of the converter process and the process risks introduced by continuing strong control under unreliable prediction.

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and do not constitute a limitation on the scope of protection of this invention.

[0064] The core steps and principles of the online prediction and control method for converter endpoints based on multimodal data and symbolic regression, as described in the preceding text, have been systematically explained. To verify the effectiveness of this invention, demonstrate its specific application process, and quantify its technical effects, three specific implementation examples will be provided below, based on actual production data from a 300-ton converter in a steel plant. These examples are derived from real experimental records and effect verification during the research and development process, fully presenting the entire process from data acquisition, model construction, online prediction to operation guidance, and providing key performance indicators.

[0065] Example 1: Model Building and Basic Performance Validation. This example mainly demonstrates the process of training a symbolic regression prediction model using historical data, and provides the model's basic mathematical expression and prediction accuracy on the test set.

[0066] First, historical data from 1000 furnaces were collected to train the YOLOv8 segmentation model and the symbolic regression model. A 2-megapixel high-speed industrial camera and an ultraviolet-visible spectrometer (200–800 nm) were used.

[0067] Implementation process: Real-time acquisition of furnace flame images and spectral data; real-time segmentation of flame regions using the YOLOv8 model with an accuracy of 98.2%; extraction of 25-dimensional visual features and 15-dimensional spectral features; and construction of a 60-dimensional fusion feature vector by combining 20-dimensional process parameters.

[0068] The carbon content prediction formula generated by symbolic regression is: , The temperature prediction formula generated by symbolic regression is: .

[0069] The online prediction accuracy for carbon content (error ±0.015%) reached 92.3%, and the accuracy for temperature (error ±12℃) reached 94.1%.

[0070] Example 2: Parameter Weighting Analysis and Operational Guidance. This example demonstrates how to utilize the interpretability of a symbolic regression model to perform parameter sensitivity analysis and generate specific operational guidance suggestions, based on a carbon content prediction model.

[0071] In the converter where the method of this invention is implemented, the carbon content prediction formula obtained by symbolic regression is as follows: The results of the parameter weighting analysis are as follows: Oxygen blowing volume ( ): Weighting coefficient -0.002, highest sensitivity, for every 100 Nm³ increase in oxygen content, carbon content decreases by 0.02%; Iron silicon content ( ): Weighting coefficient -0.025, with a significant impact; for every 0.1% increase in silicon content, the endpoint carbon content decreases by 0.0025%; K / Na spectral ratio ( ): Sensitive within a specific range due to the nonlinear influence of the tanh function; flame brightness ( ): The change is gradual due to the influence of the logarithmic function.

[0072] Example of generating operation instructions: When the predicted carbon content is 0.06% and the target carbon content is 0.04%, the carbon deviation is -0.02%, the oxygen blowing sensitivity is -0.00002 / % (per Nm³), and it is recommended to increase the oxygen blowing by 0.02 / 0.00002=1000 Nm³.

[0073] Taking safety into account, the actual recommendation is: "Increase the oxygen supply by 800 Nm³, which is expected to reduce the carbon content to 0.041%"; at the same time, considering temperature compensation, increasing the oxygen supply by 800 Nm³ is expected to raise the temperature by about 12°C. If the temperature is already close to the upper limit, it is also recommended to: "Add 2 tons of scrap steel to control the temperature rise."

[0074] Example 3: Full-Process Online Application Instance. This example simulates a complete online prediction and control process in the later stages of refining, gradually demonstrating the complete closed loop from real-time data acquisition, feature extraction, model prediction, confidence assessment to multi-objective optimization to generate operational guidance, and finally comparing and verifying with actual measurement results.

[0075] The steel grade being smelted is Q235B, with a target carbon content of 0.05% and a target temperature of 1680℃. Currently, it is in the late stage of the blowing process (oxygen blowing time 12 minutes).

[0076] Step 1: Multimodal Data Acquisition 1.1 Process parameter acquisition: Molten iron weight: 280 tons; Molten iron temperature: 1250℃; Iron composition: C 4.2%, Si 0.45%, Mn 0.35%, P 0.12%, S 0.03%; Scrap steel addition amount: 20 tons (60% heavy scrap steel, 30% light scrap steel, 10% pig iron blocks); Oxygen blowing parameters: Oxygen blowing volume 18000 Nm³, oxygen blowing time 12 minutes, oxygen supply intensity 150 Nm³ / min; Flux added: 6000 kg of lime, 800 kg of lightly calcined dolomite; Furnace age: 1250th furnace.

[0077] 1.2 Simultaneous acquisition of visual and spectral data: Industrial camera: Captures RGB images of the furnace flame (1920×1080 resolution, 25fps). Spectrometer: Collects spectral data in the wavelength range of 200–800 nm with a resolution of 0.1 nm.

[0078] Step 2: Flame Area Location and Segmentation YOLOv8 model processing results: Input: The flame image of the current frame; Output: Flame region segmentation mask, bounding box coordinates [x:350, y:220, width:420, height:380]; Confidence level: 0.98; Processing time: 45ms; Technical details: The model is trained on 5,000 labeled images, including the flame morphology in the early, middle and late stages of the blowing process, to ensure accurate recognition throughout the entire cycle.

[0079] Step 3: Multi-dimensional feature extraction 3.1 Visual feature extraction results: Color characteristics: Hue mean = 28.5, Hue variance = 15.2, Saturation mean = 0.85, Luminance mean = 0.78, Luminance maximum = 0.92, Luminance 95th percentile = 0.88; Texture features: contrast = 0.45, energy = 0.62, correlation = 0.78, homogeneity = 0.81; Morphological characteristics: Flame area = 158400 pixels, aspect ratio = 1.105, roundness = 0.72; Dynamic characteristics: Brightness fluctuation rate = 0.15, area fluctuation rate = 0.08.

[0080] 3.2 Spectral feature extraction results: Characteristic spectral line intensities: Potassium 769nm spectral line intensity = 2850, Sodium 589nm spectral line intensity = 3200, Calcium 422nm spectral line intensity = 1250, Iron 371nm spectral line intensity = 980; Continuous spectrum characteristics: spectral centroid = 512.3 nm, spectral width = 85.6 nm; Spectral line ratios: Potassium-sodium intensity ratio = 0.891, calcium-iron intensity ratio = 1.276.

[0081] Step 4: Feature Fusion and Model Prediction 4.1 Feature Fusion: Visual features: 15 dimensions; Spectral characteristics: 8 dimensions; Process parameters: 22 dimensions; The fused feature vector has 45 dimensions. After PCA dimensionality reduction: 28 dimensions (95% variance retained).

[0082] 4.2 Symbolic Regression Model Prediction: Carbon content prediction model: , Predicted result: Carbon content = 0.067%. , Predicted result: Temperature = 1668℃.

[0083] Step 5: Parameter weighting analysis and confidence assessment 5.1 Sensitivity Analysis (Under Current Operating Conditions): Carbon content sensitivity (% / unit change) (For every 100 Nm³ increase in oxygen, carbon decreases by 0.018%) (For every 0.1 increase in the K / Na ratio, carbon content decreases by 0.0012%) (Brightness increased by 10%, carbon increased by 0.0008%) Temperature sensitivity (°C / unit change) (For every 100 Nm³ of oxygen added, the temperature increases by 1.3 °C). (For every additional ton of scrap steel, the temperature decreases by 0.35℃). (For every 0.1% increase in silicon content, the temperature increases by 2.5°C).

[0084] 5.2 Confidence assessment: Current performance of the sliding window (50 furnaces): Carbon content prediction , Temperature forecast , Confidence calculation: , , The average confidence level is 0.915, which is greater than 0.9 and falls within the normal operating range.

[0085] Step 6: Generating and Executing Operation Instructions 6.1 Target Deviation Analysis: Carbon content deviation: (Slightly high) Temperature deviation: (Too low).

[0086] 6.2 Multi-objective optimization solution: Optimization issues: , Constraints: , , Solution results: Optimal adjustment amount: , .

[0087] 6.3 System Output Operation Guidelines: Current status: Carbon content is too high (0.067%), temperature is too low (1668℃); Recommended action: Increase oxygen blowing volume by 220 Nm³; Expected results: The carbon content decreased to 0.049% (±0.005); The temperature rose to 1675℃ (±10℃); Double hit probability: 89%; Confidence level: 91.5%, can be executed automatically.

[0088] 6.4 Actual Implementation and Verification: The operator confirmed and executed the oxygenation at 220 Nm³. Continue blowing for 1.5 minutes, then stop blowing; Actual measurement results: carbon content 0.048%, temperature 1678℃; Both hits were successful, and the deviations were within the acceptable range.

[0089] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for online prediction of converter endpoint carbon content and temperature based on multimodal data, characterized in that, The method includes: S1. Collect visual image data of the furnace mouth flame, spectral data of the furnace mouth flame, and process parameter data of the converter blowing process to form multimodal data. S2. Input the visual image data into the target detection and segmentation model to locate and segment the flame area at the furnace opening, and obtain the spatial location information of the flame area. S3. Based on the spatial location information, extract visual features from the segmented flame region, and extract spectral features corresponding to the space of the flame region from the spectral data; S4. The visual features, the spectral features, and the process parameter data are fused and preprocessed to construct a fused feature vector; S5. Obtain a pre-trained symbolic regression prediction model, input the fused feature vector into the symbolic regression prediction model, and obtain the predicted values ​​of the endpoint carbon content and the endpoint temperature. S6. Based on the deviation between the predicted value and the target value, and the sensitivity of each input parameter under the current operating conditions, generate online operation guidance information for adjusting process parameters; S7. Collect actual endpoint test values, update the coefficients of the symbolic regression prediction model online in a sliding window manner, and switch to conservative control mode when the confidence level is lower than the preset confidence threshold.

2. The method according to claim 1, characterized in that, The process parameter data includes at least the weight of molten iron, the temperature of molten iron, the chemical composition of molten iron, the amount of scrap steel added, the amount of flux added, the total amount of oxygen blown, the oxygen blowing intensity, the oxygen blowing time, the oxygen lance height, the bottom blowing gas flow rate, the current blowing time, and the furnace age.

3. The method according to claim 1, characterized in that, The target detection and segmentation model is a deep learning-based flame segmentation model that outputs pixel-level segmentation masks and / or bounding box coordinates of the flame region. The target detection and segmentation model is trained using a dataset of flame images covering the early, middle, severe decarburization, and final stages of converter blowing.

4. The method according to claim 1, characterized in that, The visual features include at least three of the following: color features, texture features, morphological features, and dynamic change features of the flame region.

5. The method according to claim 4, characterized in that, The color features include statistical features of the H, S, and V components extracted from the HSV color space; the texture features include texture statistical features extracted based on the gray-level co-occurrence matrix; the morphological features include the area, aspect ratio, and roundness of the flame region; and the dynamic change features include flame brightness fluctuation features and / or flame region area fluctuation features.

6. The method according to claim 1, characterized in that, The spectral features include characteristic element spectral line intensity features, continuous spectrum features, and spectral line ratio features. The characteristic element spectral line intensity features include the integrated intensity of spectral lines of at least two elements selected from potassium, sodium, calcium, and iron. The continuous spectrum features include the spectral centroid and spectral width. The spectral line ratio features include the potassium-sodium intensity ratio and / or the calcium-iron intensity ratio.

7. The method according to claim 1, characterized in that, The fusion and preprocessing in S4 includes standardizing and / or normalizing the fused features, and using dimensionality reduction methods to reduce the dimensionality of the fused feature vectors to reduce feature correlation.

8. The method according to claim 1, characterized in that, The symbolic regression prediction model is trained using a symbolic regression method based on genetic programming. During the training process, a multi-objective optimization strategy is employed to minimize both prediction error and expression complexity. A penalty term is also applied to the expression complexity to obtain a prediction model with a concise structure.

9. The method according to claim 8, characterized in that, When training the symbolic regression prediction model, metallurgical process constraints are introduced, which include at least directional constraints and boundary constraints. The directional constraints are used to ensure that the direction of influence of variables related to oxygen blowing on the endpoint carbon content conforms to the laws of metallurgical physics, and the boundary constraints are used to limit the predicted values ​​of the endpoint carbon content and / or the predicted values ​​of the endpoint temperature within a preset physical reasonable range.

10. The method according to claim 1, characterized in that, The sensitivity is obtained by taking the partial derivative of the symbolic regression prediction model with respect to each input parameter, and then standardizing the partial derivative results. The online operation guidance information determines the direction and amount of adjustment for oxygen blowing volume, coolant addition volume, and / or oxygen lance height by combining the deviations of the predicted endpoint carbon content and the predicted endpoint temperature from the target value, under the condition of meeting the process parameter adjustment range constraints.

Images