Steel ladle refining slag chemical component detection method based on machine vision

Abstract

The invention belongs to the technical field of converter steelmaking, and discloses a steel ladle refining slag chemical component detection method based on machine vision. The method aims at solving the problem of analysis lag in the prior art. In order to realize efficient and accurate identification, a slag sample data set is constructed, each group comprises slag sample photos and corresponding test components, and multi-angle and multi-illumination optimization processing is performed on image data so as to improve the generalization ability and robustness of the model. A training strategy of the model focuses on the utilization of a CNN technology, such as special processing for illumination and angle changes, so as to ensure innovativeness and high recognition precision, and training is based on a large amount of historical production data, and masters a mapping relation between visual features and chemical components of the slag sample. And collecting current production data and slag sample pictures, and inputting the current production data and the slag sample pictures into the trained model, so that automatic prediction of the final chemical components of the ladle refining slag can be realized.

Description

Technical Field

[0001] This invention relates to the field of converter steelmaking technology, and in particular to a method for detecting the chemical composition of ladle refining slag based on machine vision. Background Technology

[0002] Currently, my country's steel production mainly relies on a long-process converter smelting method. Ladle refining furnaces (such as LF furnaces and RH furnaces) play a crucial role in the steelmaking process. Their core function is to precisely adjust the composition of molten steel through the control of refining slag composition, thereby meeting the quality and performance requirements of different steel grades. The chemical composition of the refining slag, especially its oxide content (such as CaO, SiO2, Al2O3, MgO, FeO, etc.), directly affects key metallurgical processes such as desulfurization and inclusion morphology analysis.

[0003] Traditional methods for analyzing the composition of ladle refining slag primarily rely on laboratory testing, such as X-ray fluorescence spectrometry (XRF) or chemical titration. However, these methods have inherent limitations, the most prominent being time lag. The time required from collecting slag samples, sending them to the laboratory for analysis, to obtaining the final test results is often substantial. This time lag severely hinders real-time optimization and precise control of the refining process. In continuous steel production processes, especially in smelting steps with extremely stringent composition requirements, such as LF deep desulfurization or RH ultra-low carbon steel production, if the refining slag composition deviates from the target range and is not adjusted in time, it can lead to substandard steel composition, resulting in scrap steel, affecting product quality, and even causing production accidents, resulting in significant economic losses.

[0004] Therefore, there is an urgent need for a method that can identify the chemical composition of ladle refining slag in real time, efficiently, and accurately, in order to overcome the problem of existing analysis lag and empower the intelligent upgrading of the steel production process. Although there have been studies that have attempted to use image recognition technology to assist in slag composition analysis, most of them focus on specific processes or components, and when faced with the complex and varied lighting, angles, and microscopic morphological differences of the slag samples in actual production, the generalization ability and recognition accuracy of the models still have considerable room for improvement. Summary of the Invention

[0005] (I) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present invention provides a machine vision-based method for detecting the chemical composition of ladle refining slag, which solves the problem of lagging technology in the analysis of chemical composition in converter smelting.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, this invention mainly provides a machine vision-based method for detecting the chemical composition of ladle refining slag, the main steps of which include:

[0009] S1. Collect multi-view photos and historical production data of the ladle refining slag from the converter under test; the historical production data includes: mass fractions of calcium oxide, silicon dioxide, alumina, ferrous oxide, and magnesium oxide in the ladle refining slag; preprocess the multi-view photos, including size adjustment, normalization, and standardization, and extract slag surface texture features;

[0010] S2. Preprocess the multi-view photos of the current ladle refining slag to obtain a multi-view fused slag surface feature map; input the multi-view fused slag surface feature map into the trained ladle refining slag chemical composition identification and prediction model to predict the calcium oxide content, silicon dioxide content, aluminum oxide content, ferrous oxide content, and magnesium oxide content of the ladle refining slag of the converter to be tested.

[0011] The chemical composition prediction model for ladle refining slag is obtained by training a pre-constructed deep learning model based on historical production data and slag surface texture features.

[0012] The slag surface texture features include:

[0013] Texture features: The surface roughness, crystal grain distribution and stripe structure of slag are characterized by gray-level co-occurrence matrix and convolutional neural network activation map;

[0014] Three-dimensional morphological features: Generate depth maps based on stereo vision, and calculate surface normal vector distribution and roughness statistics;

[0015] The three-dimensional morphological features can reflect the macro- and micro-geometric undulations and spatial orientation information of the surface of the refining slag. The normal vector distribution entropy is used to measure the complexity of the surface morphology. The three-dimensional morphological features and texture features are jointly input into a deep learning model.

[0016] Roughness statistics, as part of the three-dimensional morphology features, are specifically used to quantify and reflect the macro- and micro-geometric undulations and spatial orientation information of the refined slag surface. They are calculated together with the surface normal vector distribution from a depth map generated based on stereo vision, where the normal vector distribution entropy further measures the complexity of the surface morphology. These features are designed as inputs into a deep learning model, used in conjunction with texture features, to achieve comprehensive analysis and modeling of surface properties. These statistics help capture surface roughness-related indices, thereby improving the model's ability to understand and process material surface morphology.

[0017] Following S2, it also includes:

[0018] Based on the predicted chemical composition of the refining slag, the current production data is adjusted until the predicted chemical composition of the refining slag meets the steel alloy composition control target.

[0019] The training process for the ladle refining slag chemical composition prediction model is as follows:

[0020] A1. Collect historical production data of the converter under test and multi-view photos of the refining slag in the ladle during historical production; the historical production data is a complete dataset for each single production, including the mass fraction of calcium oxide, silica, alumina, ferrous oxide, and magnesium oxide in the refining slag.

[0021] A2. Perform data preprocessing on the historical production data, including cleaning missing and outlier values; normalizing or standardizing numerical features to eliminate the influence of different units on model training; ensuring the integrity and consistency of input data to improve the stability and prediction accuracy of model training; invalid feature values ​​include: smelting number, gun number, gun age, production date, and steel code; replace outliers in the cleaned historical production data; the outliers include: zero values, null values, and discrete values. Use the average value of historical production data in the same category as the outlier as the replacement value to replace the outlier.

[0022] A3. Based on historical production data that has undergone data preprocessing, select the calcium oxide content, silicon dioxide content, aluminum oxide content, magnesium oxide content, and ferrous oxide content corresponding to each single production run, detected by spectral analysis or XRD method.

[0023] A4. Training samples are constructed by pairing the filtered and preprocessed single-production feature data, corresponding multi-view images of ladle refining slag, and their chemical composition labels. These training samples are input into a pre-built deep learning model in a pairing format of "multi-view image set and corresponding chemical composition." The multi-view image set includes viewpoint pose and lighting condition data. The deep learning model includes: a shared-weight convolutional neural network module for extracting texture features from multi-view images; a cross-view self-attention feature fusion module for achieving cross-view information interaction; and a multimodal regression module for jointly predicting chemical composition using geometric features and texture features. During training, grouped cross-validation is used to prevent data leakage, and a weighted loss function is used to handle data imbalance. This yields a ladle refining slag chemical composition prediction model with multi-view feature fusion and chemical composition prediction capabilities.

[0024] The process of acquiring the multi-view photographs of the historical steel ladle refining slag included:

[0025] A set of multi-view images of refining slag is acquired simultaneously using at least two synchronously triggered industrial cameras. Each camera is positioned in a fixed orientation and covers key areas of the slag surface. The industrial cameras are calibrated for intrinsic and extrinsic parameters, and distortion correction is performed on the acquired images. Illumination normalization, color temperature correction, and interference processing are performed, including color constancy, dynamic range synthesis, highlight suppression, saliency separation, masking, and saturation region repair, to obtain effective texture regions. During the acquisition process, the illumination parameters of the light source are controlled in real time through the calculation and feedback of image quality evaluation indicators to adaptively optimize the imaging quality. The image quality evaluation indicators include image contrast, brightness uniformity, and signal-to-noise ratio.

[0026] The image contrast threshold is >0.5; the variance of the brightness uniformity is <10%; and the signal-to-noise ratio is >30dB.

[0027] Step A4 includes:

[0028] A41. Filter and label the historical production data to obtain the feature data and corresponding tags for each single production run;

[0029] A42. The multi-view photographs of the steel ladle refining slag from each single production and the corresponding chemical composition content of the refining slag are used as sample data, and divided into a training set and a test set; the test set is used for model testing.

[0030] A43. Optimize the hyperparameters of the prediction model, including the learning rate and batch size;

[0031] A44. The prediction model is trained end-to-end using the training set. First, multi-view images of steel ladle refining slag collected during historical production processes are input into a shared-weight convolutional neural network to extract texture features from each viewpoint. Then, feature fusion is achieved through a cross-view self-attention feature fusion module, which is invariant to the arrangement of viewpoint inputs and can adapt to changes in the number of viewpoints. Depth maps, normal maps, and roughness distributions are obtained based on multi-view image reconstruction to extract geometric features. Finally, geometric features and texture features are fused at the feature layer to form a joint representation for predicting the chemical composition of refining slag.

[0032] Following A44, it also includes:

[0033] The refining slag photographs from each single production run in the test set are input into the prediction model to generate predicted calcium oxide content, silicon dioxide content, aluminum oxide content, magnesium oxide content, and ferrous oxide content; based on the predicted and actual values, the coefficient of determination R² and root mean square error RMSE are calculated to evaluate the prediction model.

[0034] The prediction model for the chemical composition of the refining slag was tested based on the coefficient of determination and the root mean square error; the formula for calculating the coefficient of determination R² is:

[0035] The formula for calculating the root mean square error (RMSE) is as follows:

[0036] N is the total number of samples in the test set, f i The predicted value output by the model, y i For the actual values ​​of the feature data samples, The mean of the feature data samples.

[0037] The categories of characteristic data for each production run include:

[0038] Melting number, lance number, lance age, production date, steel quality code, current production refining slag photo of the converter to be tested, refining slag calcium oxide mass fraction, refining slag silica mass fraction, refining slag alumina mass fraction, refining slag ferrous oxide mass fraction, refining slag magnesium oxide mass fraction.

[0039] (III) Beneficial Effects

[0040] This invention provides a machine vision-based method for detecting the chemical composition of ladle refining slag. The method inputs the current production data of the converter under test into a trained ladle refining slag chemical composition recognition and prediction model to predict the chemical composition content. Then, based on the chemical composition content, the amount of alloy added is calculated to achieve precise control of alloy feeding, thereby achieving precise control of the alloy composition of molten steel, improving production efficiency and economic benefits.

[0041] This method, through intelligent and high-precision slag composition identification, significantly empowers intelligent steel production, and is expected to substantially improve product quality, production efficiency, and reduce production costs. For example, it can achieve FeO levels far below 1% in LF deep desulfurization steel grades, or precisely control FeO between 3% and 6% in RH ultra-low carbon steel grades. Attached Figure Description

[0042] Figure 1 This is a schematic flowchart of a machine vision-based method for detecting the chemical composition of ladle refining slag, provided in an embodiment of the present invention. Detailed Implementation

[0043] To better explain and facilitate understanding of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present invention can be understood more clearly and thoroughly, and that the scope of the present invention can be fully conveyed to those skilled in the art.

[0045] Converter steelmaking is a major long-process steelmaking method. It involves using molten iron and scrap steel, without external energy input, relying on the physical heat of the molten iron itself and the heat generated by chemical reactions between its components to complete the steelmaking process in a converter. It offers advantages such as high production speed, large output, high single-furnace output, low cost, and low investment.

[0046] With increasing market demands for steel quality and performance, the precision required for controlling the alloy content in steel is also rising. Alloying at the converter tapping stage is a crucial step in alloy content control; however, due to the lag in chemical composition analysis, the final chemical composition at the smelting endpoint is unknown when calculating alloy addition amounts. This, to some extent, restricts the precise control of alloy composition in molten steel, impacting steel quality, production efficiency, and economic benefits. Refining slag, a product generated during steelmaking, has a significant impact on production efficiency and final product quality due to its composition and properties. Therefore, its content control is particularly important. To address this, this invention proposes a machine vision-based method for detecting the chemical composition of ladle refining slag. Figure 1 In the illustrated embodiment, the specific implementation is as follows:

[0047] S1. Collect multi-view photos and historical production data of the ladle refining slag from the converter under test; the historical production data includes: mass fractions of calcium oxide, silicon dioxide, alumina, ferrous oxide, and magnesium oxide in the ladle refining slag; preprocess the multi-view photos, including size adjustment, normalization, and standardization, and extract slag surface texture features;

[0048] S2. Preprocess the multi-view photos of the current ladle refining slag to obtain a multi-view fused slag surface feature map; input the multi-view fused slag surface feature map into the trained ladle refining slag chemical composition identification and prediction model to predict the calcium oxide content, silicon dioxide content, aluminum oxide content, ferrous oxide content, and magnesium oxide content of the ladle refining slag of the converter to be tested.

[0049] The chemical composition prediction model for ladle refining slag is obtained by training a pre-constructed deep learning model based on historical production data and slag surface texture features.

[0050] In one embodiment, the historical production data may include:

[0051] Mass fractions of calcium oxide, silica, alumina, ferrous oxide, and magnesium oxide in refining slag;

[0052] In one embodiment, before S1, the method further includes training a pre-constructed initial refining slag chemical composition prediction model based on the historical production data of the converter under test, to obtain the refining slag chemical composition prediction model, including:

[0053] A1. Collect the historical production data of the converter under test. The historical production data is a complete production dataset for each single production run of the converter under test.

[0054] Collecting complete production data can avoid collecting data from the same production time multiple times in historical production data, and it can also ensure the authenticity of the data, avoiding fabricated false data or guidance figures given by human experience, thus avoiding inaccurate data samples.

[0055] In this embodiment, the categories of complete production data for each single production run include:

[0056] Photos of the current production refining slag of the converter under test, the mass fraction of calcium oxide in the refining slag, the mass fraction of silica in the refining slag, the mass fraction of alumina in the refining slag, the mass fraction of ferrous oxide in the refining slag, and the mass fraction of magnesium oxide in the refining slag.

[0057] Of course, in actual production, the categories of the above complete production data can be added or deleted on a small scale according to the actual situation.

[0058] Furthermore, after collecting historical production data from the converter under test, there may be data deviations due to operational errors by on-site data collection personnel, equipment malfunctions, or other reasons. Data from some furnaces may contain missing values, zero values, or a large degree of dispersion. Missing values ​​reduce sample information, too many outliers can bias the model's prediction results, and duplicate values ​​can reduce the variance and alter the distribution of the data. Therefore, step A2 can be implemented:

[0059] The historical production data is preprocessed, and missing values, duplicate values, and outliers are handled by deletion, replacement, interpolation, and other methods.

[0060] For example, in one embodiment, A2 includes the following implementation steps:

[0061] A21. Clean the historical production data and use metallurgical mechanism analysis to remove obvious invalid feature values:

[0062] The invalid feature values ​​may include features unrelated to the chemical composition content of the refining slag, such as smelting number, lance number, lance age, production date, and steel quality code.

[0063] A22. Replace outliers in the cleaned historical production data;

[0064] The outliers include: zero, null, and discrete values.

[0065] Specifically, A22 can be used to replace the outlier by taking the average value of historical production data that belongs to the same category as the outlier in the historical production data.

[0066] For example, in one embodiment,

[0067] The mass fraction of each chemical component in the refining slag is closely related to the content of the chemical components, so the average value is used instead.

[0068] Further, in step A3, based on the historical production data after data preprocessing, the chemical composition content of the refining slag corresponding to each single production run of the converter is selected.

[0069] In some embodiments, this step may also be referred to as establishing a mechanistic model of the converter to be tested for converter steelmaking.

[0070] Specifically, this can be implemented as follows:

[0071] A31. Based on the historical production data, the last test, i.e., the mass fraction of each chemical component in the refining slag measured by TSC, is selected as the content of each chemical component in the refining slag at the converter endpoint.

[0072] Based on the above steps, the chemical composition content of the refining slag corresponding to the complete production data of each single production run is selected. Each complete production run and its corresponding chemical composition content of the refining slag constitute a sample data set. In practical applications, as many complete production runs as possible should be collected and the corresponding chemical composition content of the refining slag calculated to avoid overfitting during model training.

[0073] For example, in one embodiment, complete data from over a thousand production runs were collected.

[0074] Further, in step A4, using metallurgical principles, the characteristic data affecting the content of each chemical component in the refining slag in the historical production data are determined.

[0075] In this embodiment, the Pearson correlation coefficient method is used for judgment and selection. The predefined correlation calculation formula is as follows:

[0076] n is the total number of furnace runs in historical production data, and xi is the actual value of the characteristic data. It is the mean of the feature data, y iThese are the actual values ​​of the content of each chemical component in the refining slag. It is the average content of each chemical component in the refining slag.

[0077] By performing correlation screening on various feature data of historical production data, we can select feature groups with strong explanatory power for the target variable from all features. That is, we can select features that are highly correlated with the content of various chemical components in refining slag from all features, thereby improving the model training effect and increasing the accuracy of predicting the content of various chemical components in refining slag.

[0078] In some other embodiments, other correlation calculation methods may be selected for calculation, which are not considered as limitations.

[0079] In the embodiments, all features include at least: a photograph of the refining slag of the converter under test, the mass fraction of calcium oxide in the refining slag, the mass fraction of silica in the refining slag, the mass fraction of alumina in the refining slag, the mass fraction of ferrous oxide in the refining slag, and the mass fraction of magnesium oxide in the refining slag.

[0080] Further, in step A4, the historical production data is filtered according to the relevant feature types, and the relevant feature data in each production data is retained. The converter refining slag photos of each production and the content of each chemical component in the refining slag of the current production are input into the pre-constructed deep learning model for training, and the resulting intelligent identification and prediction model of the chemical composition of ladle refining slag is obtained.

[0081] In practical applications, when establishing an intelligent identification and prediction model for the chemical composition of ladle refining slag based on machine learning, at least two models should be established for comparison, including but not limited to: VGGnet, ResNet, and Vision Transformer, to prevent the generated prediction model for the chemical composition of refining slag from deviating.

[0082] Specifically, in one embodiment, the initial refining slag chemical composition prediction model is the VisionTransformer refining slag chemical composition prediction model (Vision Transformer), which may include:

[0083] A41. Filter and label the historical production data to obtain the characteristic data of each single production run and the corresponding tag; specifically, this can be:

[0084] Use `data.head()` to view the new DataFrame and ensure that the features and labels are within the specified range; use `data.dropna()` to ensure that there are no NaN values ​​in the data; use `df.info()` to view the data type and data volume, and you may need to convert the data type as needed; use `df.describe()` to see the minimum, maximum, mean, median, standard deviation, and quartile range of each column.

[0085] A42. Take the photos of each batch of refined slag produced and the content of each chemical component of the refined slag produced in that batch as sample data, and input them into the Vision Transformer ladle refined slag chemical composition prediction model; divide the sample data into a training set and a test set; the test set is used to test the ladle refined slag chemical composition prediction model.

[0086] In one embodiment, the sample data is split using the train_test_split function of Scikit-learn.

[0087] A43. Optimize the hyperparameters of the prediction model for the chemical composition of ladle refining slag;

[0088] A44. Train the ladle refining slag chemical composition prediction model using the training set to obtain the ladle refining slag chemical composition prediction model.

[0089] Furthermore, in some embodiments, the following is also implemented: the characteristic data of each single production in the test set is input into the ladle refining slag chemical composition prediction model to generate the predicted content of each chemical component in the refining slag for that production.

[0090] Based on the predicted chemical composition content of the refining slag in the current production and the chemical composition content of the refining slag in the same single production in the test set, the coefficient of determination and root mean square error of the predicted chemical composition content of the refining slag in the current production are calculated.

[0091] The effectiveness of the refining slag chemical composition prediction model is determined by testing the coefficient of determination and the root mean square error.

[0092] The coefficient of determination (R²) represents the goodness of fit of the model; the root mean square error (RMSE) is used as a standard to measure the prediction results of deep learning models. The root mean square error is also called the standard error. It is the arithmetic square root of the mean square error. The reason for introducing the root mean square error is exactly the same as the reason for introducing the standard deviation. That is, the dimension of the mean square error is different from the dimension of the data and cannot intuitively reflect the degree of dispersion. Therefore, the square root of the mean square error is taken to obtain the root mean square error.

[0093] The formula for calculating the coefficient of determination R² is as follows:

[0094] The formula for calculating the root mean square error (RMSE) is as follows:

[0095] N is the total number of feature data samples in the test set, f i The predicted value output by the model, y i For the actual values ​​of the feature data samples, The mean of the feature data samples.

[0096] The methods for predicting the chemical composition of refining slag provided in the above embodiments predict the content of refining slag and its chemical composition by inputting the current production data of the converter under test into a trained model. Then, based on the chemical composition content, the amount of alloy added is calculated, achieving precise control of alloy feeding and consequently, precise control of the alloy composition of molten steel, thus improving production efficiency and economic benefits. These methods predict the chemical composition of alloyed steel from the start of the converter steelmaking process to the point of tapping with high accuracy, precisely calculating the amount of alloy added and achieving precise control of alloy feeding, thereby improving production efficiency and economic benefits. This also increases the quality pass rate of steel produced from converter steelmaking, reduces the probability of remelting, and minimizes resource waste.

[0097] By collecting complete production data, model prediction biases caused by data duplication or errors in training samples are avoided. All features are screened for relevance, resulting in stronger generalization ability. This avoids excessive features and overfitting, reduces computational complexity and training difficulty, and increases model accuracy.

[0098] The model is then tested using the coefficient of determination and root mean square error to determine its accuracy.

[0099] To better explain the technical solution proposed in this invention, more specific embodiments will be used for explanation and description below.

[0100] Example 1

[0101] The converter to be tested is a 200t converter in a steel plant.

[0102] First, a model for identifying and predicting the chemical composition of the ladle refining slag in this converter is established.

[0103] Step A1: Collect historical production data of the 200t converter. The data is the actual on-site data recorded by the steel plant. After preliminary screening and elimination of incomplete data, a total of 3,000 heats of production data were collected as historical production data. Data categories (also known as data characteristics): Collect current production refining slag photos of the converter under test, refining slag calcium oxide mass fraction, refining slag silica mass fraction, refining slag alumina mass fraction, refining slag ferrous oxide mass fraction, and refining slag magnesium oxide mass fraction.

[0104] Further proceed to step A2: preprocess the collected historical data. Remove features unrelated to the chemical composition of the refining slag, such as: smelting number, lance number, lance age, production date, steel code, etc.; according to data cleaning rules, use a replacement method to process null values, zero values, and highly dispersed outlier data.

[0105] Further, in step A3, the last test of the secondary lance, namely the mass fraction of chemical components of the refining slag detected by TSO, is selected as the chemical component content of the converter refining slag. The normality test is performed on each input feature data. If the normality property is not met, the data set containing the discrete value is directly deleted.

[0106] Based on the chemical composition content of the refining slag and the historical production data, step A4 feature selection is performed to determine the feature data that has a significant impact on the chemical composition content of the ladle refining slag, i.e., the relevant feature data.

[0107] In this embodiment, the final model input variables are determined as follows: current production refining slag photos of the converter under test, the mass fraction of calcium oxide in the refining slag, the mass fraction of silica in the refining slag, the mass fraction of alumina in the refining slag, the mass fraction of ferrous oxide in the refining slag, and the mass fraction of magnesium oxide in the refining slag. Relevant refining slag photos from the 1500 heats are selected, and the relevant feature data are divided into a training set and a test set with a ratio of 8:2. In this embodiment, a total of 1500 heats are selected as sample data, with 1200 sets of data randomly selected as the training set and 300 sets of data selected as the test set.

[0108] The feature data is input into a pre-constructed intelligent identification and prediction model for the chemical composition of refining slag for training. The hyperparameter tuning specifically includes: Batch Size: Batch size, defined as Batch Size = 32; learning_rate: Learning rate, defined as learning_rate = 0.001; Dropout: Randomly shuts down neurons during training to control overfitting, defined as Dropout = 0.5; Inputshape: Image size, defined as Inputshape = 224, 224; Epochs: Training epochs, defined as Epochs = 30.

[0109] Furthermore, in order to ensure that the model has accurate learning and prediction capabilities for training data and good prediction capabilities for new data during the learning process, this embodiment uses a selection coefficient of determination (R²). 2 The evaluation model includes the root mean square error (RMSE). Table 1 shows the evaluation indicators for the prediction model of chemical composition of refining slag from a 200t Vision Transformer converter.

[0110] Table 1:

[0111]

[0112] As shown in Table 1, within an error range of ±0.15, the prediction accuracy of chemical composition content in refining slag can reach 88.5%.

[0113] In this embodiment of the invention, the R-value of the Vision Transformer refining slag chemical composition prediction model is... 2 All values ​​reached above 0.67 and had low RMSE; the predicted values ​​fit the actual values ​​as expected, and the chemical composition content of the refining slag was within ±0.15, with a hit rate of over 85%. Therefore, the predicted values ​​of the Vision Transformer refining slag chemical composition content prediction model are very close to the actual values.

[0114] This invention provides a machine vision-based method for detecting the chemical composition of ladle refining slag. First, based on metallurgical principles and historical converter smelting data, a predictive model for the chemical composition of ladle refining slag is established. This model can directly calculate the chemical composition content of the ladle refining slag during the complex physicochemical reactions of converter smelting, thereby ensuring accurate calculation and control of alloy materials during converter tapping. Verification shows that the ladle refining slag chemical composition prediction model of this invention can accurately predict the chemical composition content of refining slag during the converter process, thus enabling more precise control of the charging process.

[0115] In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0116] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0117] In the description of this specification, the terms "one embodiment," "some embodiments," "embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0118] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make modifications, alterations, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A machine vision-based method for detecting the chemical composition of ladle refining slag, characterized in that, include: S1. Collect multi-angle photos of historical production ladle refining slag and historical production data of the converter under test; The historical production data includes: mass fractions of calcium oxide, silicon dioxide, aluminum oxide, ferrous oxide, and magnesium oxide in the ladle refining slag; the multi-view photos are preprocessed, including size adjustment, normalization, and standardization, and slag surface texture features are extracted; S2. Preprocess the multi-view photos of the current ladle refining slag to obtain a multi-view fused slag surface feature map; input the multi-view fused slag surface feature map into the trained ladle refining slag chemical composition identification and prediction model to predict the calcium oxide content, silicon dioxide content, aluminum oxide content, ferrous oxide content, and magnesium oxide content of the ladle refining slag of the converter to be tested. The chemical composition prediction model for ladle refining slag is obtained by training a pre-constructed deep learning model based on historical production data and slag surface texture features.

2. The intelligent detection method for chemical composition of ladle refining slag based on machine vision according to claim 1, characterized in that, The slag surface texture features include: Texture features: The surface roughness, crystal grain distribution and stripe structure of slag are characterized by gray-level co-occurrence matrix and convolutional neural network activation map; Three-dimensional morphological features: Generate depth maps based on stereo vision, and calculate surface normal vector distribution and roughness statistics; Roughness statistics are part of the three-dimensional morphology features, specifically used to quantify and reflect the macro- and micro-geometric undulations and spatial orientation information of the refining slag surface; the roughness statistics are calculated together with the surface normal vector distribution from the depth map generated based on stereo vision, wherein the normal vector distribution entropy is further used to measure the complexity of the surface morphology. This 3D shape feature and texture feature are jointly input into a deep learning model.

3. The method for detecting the chemical composition of ladle refining slag based on machine vision according to claim 1, characterized in that, Following S2, it also includes: Based on the predicted chemical composition of the refining slag, the current production data is adjusted until the predicted chemical composition of the refining slag meets the steel alloy composition control target.

4. The method for detecting the chemical composition of ladle slag based on machine vision according to claim 1, characterized in that, The training process for the ladle refining slag chemical composition prediction model is as follows: A1. Collect historical production data of the converter under test and multi-view photos of the refining slag in the ladle during historical production; the historical production data is a complete dataset for each single production, including the mass fraction of calcium oxide, silica, alumina, ferrous oxide, and magnesium oxide in the refining slag. A2. Perform data preprocessing on the historical production data, including cleaning missing and outlier values; normalize or standardize numerical features to eliminate the influence of different units on model training; ensure the integrity and consistency of input data, and improve the stability and prediction accuracy of model training. A3. Based on historical production data that has undergone data preprocessing, select the calcium oxide content, silicon dioxide content, aluminum oxide content, magnesium oxide content, and ferrous oxide content corresponding to each single production run, detected by spectral analysis or XRD method. A4. Pair the filtered and preprocessed single production feature data, the corresponding multi-view images of ladle refining slag, and their chemical composition labels to construct training samples; the training samples are input into the pre-constructed deep learning model in the form of "multi-view image set and corresponding chemical composition" pairing; the multi-view image set includes viewpoint pose and lighting condition data; The deep learning model includes: a shared-weight convolutional neural network module for extracting texture features from multi-view images; a cross-view self-attention feature fusion module for realizing cross-view information interaction; and a multimodal regression module for predicting chemical composition by combining geometric features and texture features. During training, a grouped cross-validation method is used to prevent data leakage, and a weighted loss function is used to handle data imbalance. A chemical composition prediction model for ladle refining slag with multi-view feature fusion and chemical composition prediction capabilities is obtained.

5. The method for detecting the chemical composition of refining slag based on machine vision according to claim 4, characterized in that, The process of acquiring the multi-view photographs of the historical steel ladle refining slag included: A set of multi-view images of refining slag is acquired simultaneously using at least two synchronously triggered industrial cameras. Each camera is positioned in a fixed orientation and covers key areas of the slag surface. The industrial cameras are calibrated for intrinsic and extrinsic parameters, and distortion correction is performed on the acquired images. Illumination normalization, color temperature correction, and interference processing are performed, including color constancy, dynamic range synthesis, highlight suppression, saliency separation, masking, and saturation region repair, to obtain effective texture regions. During the acquisition process, the illumination parameters of the light source are controlled in real time through the calculation and feedback of image quality evaluation indicators to adaptively optimize the imaging quality. The image quality evaluation indicators include image contrast, brightness uniformity, and signal-to-noise ratio.

6. The method for detecting chemical composition of refining slag based on machine vision according to claim 5, characterized in that, The image contrast threshold is >0.5; the variance of the brightness uniformity is <10%; and the signal-to-noise ratio is >30dB.

7. The method for detecting the chemical composition of refining slag based on machine vision according to claim 4, characterized in that, Step A4 includes: A41. Filter and label the historical production data to obtain the feature data and corresponding tags for each single production run; A42. The multi-view photographs of the steel ladle refining slag from each single production and the corresponding chemical composition content of the refining slag are used as sample data, and divided into a training set and a test set; the test set is used for model testing. A43. Optimize the hyperparameters of the prediction model, including the learning rate and batch size; A44. The prediction model is trained end-to-end using the training set. First, multi-view images of steel ladle refining slag collected during historical production processes are input into a shared-weight convolutional neural network to extract texture features from each viewpoint. Then, feature fusion is achieved through a cross-view self-attention feature fusion module, which is invariant to the arrangement of viewpoint inputs and can adapt to changes in the number of viewpoints. Depth maps, normal maps, and roughness distributions are obtained based on multi-view image reconstruction to extract geometric features. Finally, geometric features and texture features are fused at the feature layer to form a joint representation for predicting the chemical composition of refining slag.

8. The method for detecting the chemical composition of refining slag based on machine vision according to claim 7, characterized in that, Following A44, it also includes: The refining slag photographs from each single production run in the test set are input into the prediction model to generate predicted calcium oxide content, silicon dioxide content, aluminum oxide content, magnesium oxide content, and ferrous oxide content; based on the predicted and actual values, the coefficient of determination R² and root mean square error RMSE are calculated to evaluate the prediction model. The prediction model for the chemical composition of the refining slag was tested based on the coefficient of determination and the root mean square error; the formula for calculating the coefficient of determination R² is: ； The formula for calculating the root mean square error (RMSE) is as follows: ； N is the total number of samples in the test set, f i y is the predicted value output by the model. i For the actual values ​​of the feature data samples, The mean of the feature data samples.

Images