Graph model flame image feature extraction-based prediction method for carbon content at converter steelmaking endpoint

Abstract

The application discloses a converter steelmaking end-point carbon content prediction method based on flame image feature extraction of a graph model, and relates to the technical field of converter flame end-point carbon content. In order to solve the difficulty in extracting multi-direction and multi-scale irregular features of color texture of end-point converter flame image due to complex flame changes, and to solve the problem that it is difficult to distinguish the flame images with high similarity and similar carbon content corresponding to different end-point carbon contents, the application first maps the flame image in HSI space to phase space to enhance the spatial position correlation information; secondly, a derivative relationship weight formula reflecting the continuous change between vertices of different scales is given based on a complex network, and a multi-scale irregular directional weighted color texture complex network of the converter flame image is constructed in combination with the direction information; finally, the vertex directional weighted degree feature is quantized to the complex network topological connection mode, the DDMCN feature of the converter flame is constructed, and a KNN regression model is established to predict the end-point carbon content.

Description

Technical Field

[0001] This invention belongs to the technical field of converter flame endpoint carbon content, specifically involving a method for predicting the endpoint carbon content of converter steelmaking by extracting features from graphical flame images. Background Technology

[0002] A crucial operation in the later stages of converter steelmaking is endpoint control, which involves removing most impurities at the end of oxygen blowing, ensuring that the carbon content and various metallic element content of the molten steel meet the requirements for tapped steel quality. However, since most impurities such as sulfur and phosphorus are removed from the molten steel in the early stages of blowing, and certain specific alloying components can be addressed after converter steelmaking, the challenge lies in achieving continuous, real-time prediction of carbon content. Accurate prediction directly affects the quality of tapped steel and helps reduce energy and raw material waste.

[0003] Currently, methods for detecting the carbon content at the end point of converter steelmaking mainly include contact detection techniques, manual experience and sampling analysis, secondary lance detection, and furnace gas analysis. However, these methods have limitations such as low prediction accuracy, cost, and applicability. With the development of computer technology, non-contact detection technologies such as spectral imaging, soft measurement methods for production process data, and methods based on converter furnace mouth flame image feature extraction have emerged. Spectral imaging methods have high accuracy, but image acquisition is easily affected by on-site environmental factors. Soft measurement methods for production process data are convenient for data acquisition, but production process data is easily affected by ore quality and fluctuates significantly. Methods based on converter furnace mouth flame image feature extraction are less constrained by environmental factors and can be applied to most small and medium-sized converter steelmaking scenarios. Data acquisition is simple and convenient, with high real-time performance and low cost.

[0004] The carbon content of molten steel during smelting is closely related to flame texture and other characteristics. Existing literature shows that changes in carbon content lead to changes in the intensity of the carbon-oxygen reaction, which in turn alters the flame mechanism. The large amount of high-temperature CO gas produced will ignite immediately upon contact with air. The amount of CO emitted per unit time can be judged based on the shape, length, color, and brightness of the flame at the furnace mouth. This can also be used as a measure of the decarburization rate in the molten metal and is a major basis for judgment using empirical methods.

[0005] The key to predicting the carbon content at the end of steelmaking through converter furnace flame image feature extraction is the accurate representation of flame features, which can be mainly divided into convolutional neural network feature description and traditional feature extraction. Neural networks have the advantage of extracting deep features in flame image feature description, reducing manual intervention while offering high real-time performance, but they require a large amount of training data and high computational complexity. Traditional features include color features, region edge features, and texture features. Flame color is an intuitive piece of information that shows regular changes at different blowing stages. Existing studies have used region clustering analysis to directly extract the furnace flame color features after principal component analysis transformation, effectively reflecting different blowing stages of converter steelmaking. Prediction methods based on flame color feature extraction are simple and direct, but the flames are highly similar and the color distribution is relatively concentrated, so the extracted flame features are not sufficient. For flame region edge features, some studies have used boundary features to determine the end of converter steelmaking with high accuracy. Since flame texture can better take into account both macroscopic properties and subtle local changes, it has more obvious discriminative information compared to other flame features, therefore, many scholars have proposed methods for extracting flame texture. Some studies have used gray-level co-occurrence matrices to extract texture features from images, and then used changes in texture features to determine the steelmaking endpoint, verifying the effectiveness of flame image texture features. Other studies have further combined the multi-directional and multi-scale characteristics of flame texture with multi-trend encoding to obtain the color texture features of the furnace mouth flame, and conducted experiments to predict the carbon content at the endpoint of converter steelmaking. This verified the correlation between the directional information of flame texture and carbon content, proving its significance for the color texture of the furnace mouth flame. However, flame textures are multi-directional, multi-scale, and irregular, making traditional feature extraction algorithms insufficiently precise in their feature descriptions. The HSI (Hue-Saturation-Intensity (Lightness), HSI or HSL) color model uses three parameters, H, S, and I, to describe color characteristics. H defines the frequency of a color, called hue, which is an attribute describing a pure color; S represents the lightness or darkness of a color, called saturation, which measures the degree to which a pure color is diluted by white light; and I represents intensity or brightness, which is a subjective description. In summary, the problem with existing methods for predicting the relationship between flame textures and the carbon content of converter steelmaking is that feature acquisition is difficult due to flame variations, and flame images are similar under different carbon contents. Existing methods do not provide a precise description of the key features of flame textures and cannot accurately predict the carbon content of converter steelmaking from the irregular features of flame textures. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, this invention provides a method for predicting the carbon content at the endpoint of converter steelmaking by extracting features from graph model flame images. This method solves the difficulties in extracting multi-directional, multi-scale, and irregular color texture features from the flame images at the converter steelmaking endpoint due to the complexity of flame changes. It also addresses the problem of high similarity between flame images corresponding to different endpoint carbon contents and the difficulty in distinguishing flame images with similar carbon contents. Based on complex network texture features, this method achieves real-time and accurate prediction of carbon content.

[0007] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solution:

[0008] A method for predicting the final carbon content in converter steelmaking based on feature extraction from graphical flame images, with the following specific steps:

[0009] Step 1: Collect video footage of the flames at the furnace opening in different heats during actual converter steelmaking production, captured by industrial cameras.

[0010] Step 2: Sample the captured video frame by frame to obtain the flame image dataset, which will be used as experimental data. The label is the carbon content value of molten steel measured by the secondary lance detection technology at the end of the converter steelmaking process.

[0011] Step 3: Use a Derivative nonlinear mapping direction weighted multilayer complex network (DDMCN) to describe the color texture features of flames in the flame image dataset;

[0012] Step 4: Input the extracted color texture features into the KNN regression model to predict the endpoint carbon content, and obtain the converter endpoint carbon content prediction result through five-fold cross-validation. The five-fold cross-validation method is more accurate and stable than the existing direct validation method.

[0013] Preferably, in step 3, a complex network of directional weighted color texture features for the furnace flame image is constructed to extract multi-scale irregular directional detail information of the furnace flame, thereby obtaining texture difference features adapted to flame changes. The extraction process is as follows:

[0014] Step 3.1: After removing the furnace mouth boundary information, construct the flame image phase space mapping map;

[0015] Step 3.2: Based on complex networks, define a derivative information weighting formula that reflects the continuous changes of vertices at different scales, integrate directional information, and construct a multi-scale irregular directional weighted color texture complex network descriptor for furnace flame images. Combine the color channel information and high-order local derivatives of the flame image HSI space to construct a directional weighted multi-layer network mode; strengthen the description of color information, thereby addressing the high similarity of random natural textures presented by flame images with different carbon contents at the steelmaking endpoint, and solving the difficulty of extracting irregular features of flame textures in multiple directions and at multiple scales.

[0016] Step 3.3: Calculate the directional weighted feature quantization of the complex network topology connection pattern of the flame image to describe the vertex directional connection information, construct the color texture feature of the furnace mouth flame image, extract multi-scale irregular directional detail information of the furnace mouth flame, and obtain texture difference features that adapt to the changes of the flame.

[0017] Preferably, in step 3.1, the specific steps for constructing the flame image phase space mapping map after removing the furnace mouth boundary information are as follows:

[0018] Step 3.1.1: Remove the parts other than the flame by segmentation using the Otsu's method, and retain the target area of ​​the flame.

[0019] Step 3.1.2: Construct spatial mutual information of local textures and describe the spatial changes of flame textures in phase form. Calculate phase maps in the H, S, and I color channels respectively, and statistically analyze the structural differences in local flame regions;

[0020] Divide the image into several non-overlapping segments. The phase values ​​within the blocks are calculated, and useful information from the original image is preserved. The local flame characteristics are statistically analyzed. The flame image is described in phase form, and its mathematical expression is shown in equation (1) below:

[0021] (1)

[0022] In the formula, The average value of pixels surrounding the center pixel of a local block in a flame image is shown in Equation (2), which reduces the interference of noise in the flame image. Indicates the intensity value of surrounding pixels. The transformed phase value is represented by the spatial domain representation of the flame image, which is constructed by integrating the surrounding pixels and the average value, and thus takes into account the spatial information between pixels.

[0023] (2)

[0024] Preferably, in step 3.2, the specific steps for constructing a multi-scale, irregularly oriented, weighted color texture complex network descriptor for the furnace flame image are as follows:

[0025] Step 3.2.1: Model a network from the phase map of the flame image.

[0026] If a flame image exists Phase space map of a certain color channel conversion width and height common 1 pixel, i.e. If there are vertices, then the phase value of each pixel is used as the value of the vertex;

[0027] Step 3.2.2: Construct network connection rules by combining the derivative information, weight formula, and direction information between vertices;

[0028] Considering the derivative information between a set of vertices, the vertex pairs of the flame image phase space map complex network are... and , The connection weight between them is defined as and , The value of the intensity difference between them after derivative mapping is calculated as shown in equation (3-4):

[0029] (3)

[0030] (4)

[0031] In the formula As vertices and , The first-order derivative relationship and the second-order derivative mapping relationship between them are shown in (5-6), respectively. The connection weights represent the first-order derivative relationships between vertices. This represents the connection weights based on the second derivative relationship between vertices. The smaller the value, the closer the changing trends between vertices are, and the greater the similarity.

[0032] (5)

[0033] (6)

[0034]

[0035] In the formula This represents the phase value at the center vertex. express Phase values ​​of adjacent scale vertices, This represents the nonlinear mapping value of the first derivative after phase transformation of the flame image. and express Phase values ​​of adjacent scale vertices, This represents the transformed second derivative nonlinear mapping value. As defined in equation (7), represents the mean.

[0036] (7)

[0037] Consider the spatial orientation relationships of a set of vertices, that is, the mutual changes of the flame in different directions. If there exists a... Local region, with vertices Taking the eight-neighbor domain as an example, the direction is represented by the sequence number. The formula is expressed as (8):

[0038] (8)

[0039] calculate Serial number is Vertex in direction Pairs with neighboring points The difference, after being mapped nonlinearly by the derivative, yields the following difference pair: As shown in equation (9):

[0040] (9)

[0041] Combining information about the direction of flame changes, vertex changes are used as network connection conditions. The direction of the arrow indicates a pixel change from small to large. A parallel trend is satisfied when the center vertex and its neighboring vertices change from small to large or from large to small; a non-parallel trend is satisfied when all neighboring vertex values ​​are greater than or less than the center vertex. In other words, when the vertex... and neighboring point pairs When the change satisfies a parallel trend, then... Establish network connections; conversely, when changes satisfy a non-parallel trend, then... No network connection is established, which defines it as The connection method between the vertex and the surrounding vertices is expressed mathematically as shown in equation (10):

[0042] (10)

[0043] In the formula, A value of 1 indicates a connection, while 0 indicates no connection. .

[0044] Step 3.2.3: Construct an undirected weighted network according to the connection rules;

[0045] By combining color information, a multi-layered mapping direction weighted complex network is constructed. Each pixel in the phase spectrum of each color channel is considered a vertex, and the total number of vertices is... , forming a set , forming the edge Superimposing three color channels yields a multi-layered, direction-weighted complex network for the flame image;

[0046] Preferably, in step 3.3, the specific steps for constructing the color texture features of the furnace flame image are as follows:

[0047] Step 3.3.1: Define the vertex orientation weighting degree of the complex network for the flame image as the vertex orientation weighting degree. The formula, obtained by summing the direction parameters of the other connected vertices, is as follows:

[0048] (11)

[0049] In the formula, For the direction parameter, this paper sets... To achieve a set of parallel trend vertex direction parameters summing to 1, Used to compute each vertex of the flame image texture network The directional weighting describes the texture changes of the flame.

[0050] By computing each vertex of the texture network degree , histogram degree The calculation is shown in equation (12):

[0051] (12)

[0052] in Defined by equation (13):

[0053] (13)

[0054] Step 3.3.2: By analyzing the degree histogram of the complex network of the flame image, the features are calculated as multi-directional, multi-scale, and irregular color texture features of the flame image;

[0055] Calculate the directional weighting of complex networks The statistical features corresponding to its histogram are used as color texture descriptors for flame images. The feature vectors describing each layer of the network in the flame image are... As shown in equation (14), the flame image DDMCN color texture descriptor is constructed by fusing multi-layer network information.

[0056] (14)

[0057] in, It is a threshold. The calculation yielded the result.

[0058] (15)

[0059] The beneficial effects of this invention are:

[0060] A graph model-based method for predicting the final carbon content of converter steelmaking by extracting flame image features is proposed. First, the flame image in HSI space is mapped to phase space to enhance spatial location correlation information. Second, a derivative relationship weighting formula reflecting the continuous changes between vertices at different scales is given based on a complex network. Combined with directional information, a multi-scale, irregularly oriented, weighted color texture complex network for furnace mouth flame images is constructed to solve the problem of difficulty in obtaining flame image features that distinguish different carbon contents. Finally, the topological connection pattern of the complex network is quantified by calculating the vertex directional weighting features, constructing the DDMCN feature of the furnace mouth flame, and establishing a KNN regression model to predict the final carbon content.

[0061] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0062] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a framework diagram of the converter steelmaking endpoint carbon content prediction method based on graph model flame image feature extraction as described in this invention;

[0064] Figure 2 This refers to the process of extracting complex network color texture features from flame images in step 3 of this invention.

[0065] Figure 3 This is a scatter plot of carbon content prediction obtained by the method of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] Example 1

[0068] like Figure 1-2As shown

[0069] A method for predicting the final carbon content in converter steelmaking based on feature extraction from graphical flame images, with the following specific steps:

[0070] Step 1: Collect video footage of the flames at the furnace opening in different heats during actual converter steelmaking production, captured by industrial cameras.

[0071] Step 2: When preparing experimental data, the video is sampled frame by frame to obtain a flame image dataset, labeled as the carbon content value of molten steel measured by the secondary lance detection technology at the end of the converter steelmaking process.

[0072] Step 3: Use a derivative nonlinear mapping directional weighted multilayer complex network to describe the color texture features of flames in the flame image dataset;

[0073] Step 4: Input the extracted color texture features into the KNN regression model to predict the endpoint carbon content, and obtain a relatively accurate and stable converter endpoint carbon content prediction result through five-fold cross-validation.

[0074] In step 3, a complex network-based directional weighted color texture feature model of the furnace flame image is constructed to extract multi-scale irregular directional detail information of the furnace flame, thereby obtaining more effective texture difference features to adapt to flame changes. The extraction process is as follows:

[0075] Step 3.1: After removing the furnace mouth boundary information, construct the flame image phase space mapping map;

[0076] Step 3.1.1: Use the Otsu's method to segment and remove parts other than the flame, retaining the flame region of interest.

[0077] Step 3.1.2: Construct spatial mutual information of local textures to further describe the complex spatial variations of flame textures through phase representation. Calculate phase maps in the H, S, and I color channels to statistically analyze the structural differences in local flame regions.

[0078] like Figure 2 As shown in (a), in order to preserve the useful information of the original image regarding the statistical local flame characteristics, the image is divided into several non-overlapping segments. The flame image can be described in phase form by dividing the flame into small blocks and calculating the phase value within each block. Its mathematical expression is shown in equation (1) below:

[0079] (1)

[0080] In the formula, The average value of pixels surrounding the center pixel of a local block in a flame image is shown in Equation (2), which reduces the interference of noise in the flame image. Indicates the intensity value of surrounding pixels. The transformed phase value is represented by the spatial domain representation of the flame image, which is constructed by integrating the surrounding pixels and the average value, and thus takes into account the spatial information between pixels.

[0081] (2)

[0082] By mapping the HSI space of the flame image to the phase space using a phase transformation formula, spatial detail information and spatial location correlation information can be enhanced.

[0083] Example 2

[0084] like Figure 1-2

[0085] Step 3.2: To address the challenge of extracting irregular features from flame textures across multiple directions and scales, a derivative information weighting formula reflecting the continuous changes of vertices at different scales is defined based on a complex network. By fusing directional information, a multi-scale irregular directional weighted color texture complex network descriptor for furnace flame images is constructed. Combining the color channel information and higher-order local derivatives of the flame image's HSI space, a directional weighted multi-layer network mode is constructed to enhance the description of color information, thereby better addressing the high similarity of random natural textures presented by flame images at different carbon contents at the steelmaking endpoint.

[0086] Step 3.2.1: Model a network from the phase map of the flame image;

[0087] If a flame image exists Phase space map of a certain color channel conversion width and height common 1 pixel, i.e. If there are vertices, then the phase value of each pixel is used as the value of the vertex;

[0088] Step 3.2.2: Construct network connection rules by combining the derivative information, weight formula, and direction information between vertices;

[0089] Considering the derivative information between a set of vertices, the vertex pairs of the flame image phase space map complex network are... and , The connection weight between them is defined as and , The value of the intensity difference between them after derivative mapping is calculated as shown in equation (3-4):

[0090] (3)

[0091] (4)

[0092] In the formula As vertex and , The first-order derivative relationship and the second-order derivative mapping relationship between them are shown in (5-6), respectively. The connection weights represent the first-order derivative relationships between vertices. This represents the connection weights based on the second derivative relationships between vertices. The smaller the value, the closer the trends of change between the vertices, and the greater their similarity.

[0093] (5)

[0094] (6)

[0095]

[0096] In the formula This represents the phase value at the center vertex. express Phase values ​​of adjacent scale vertices, This represents the nonlinear mapping value of the first derivative after phase transformation of the flame image. and express Phase values ​​of adjacent scale vertices, This represents the transformed second derivative nonlinear mapping value. As defined in equation (7), represents the mean.

[0097] (7)

[0098] Consider the spatial orientation relationships of a set of vertices, that is, the mutual changes of the flame in different directions. If there exists a... Local areas, such as Figure 2 (c) shows the vertex Taking the eight-neighbor domain as an example, the direction is represented by the sequence number. The formula is expressed as (8):

[0099] (8)

[0100] calculate Serial number is Vertex in direction Pairs with neighboring points The difference, after being mapped nonlinearly by the derivative, yields the following difference pair: As shown in equation (9):

[0101] (9)

[0102] By incorporating information about the changing direction of the flames, vertex changes are used as network connection conditions. For example... Figure 2 As shown in (c), the arrows indicate pixel values ​​increasing from small to large. A parallel trend is satisfied when the center vertex and neighboring vertices change from small to large or from large to small; a non-parallel trend is satisfied when all neighboring point values ​​are greater than or less than the center point. In other words, when the vertex... and neighboring point pairs When the change satisfies a parallel trend, then... Establish network connections; conversely, when changes satisfy a non-parallel trend, then... No network connection is established, which defines it as The connection method between the vertex and the surrounding vertices is expressed mathematically as shown in equation (10):

[0103] (10)

[0104] In the formula, A value of 1 indicates a connection, while 0 indicates no connection. .

[0105] Step 3.2.3: Construct an undirected weighted network according to the connection rules;

[0106] By combining color information, a multi-layered mapping direction weighted complex network is constructed. Each pixel in the phase spectrum of each color channel is considered a vertex, and the total number of vertices is... , forming a set , forming the edge Superimposing three color channels yields a multi-layered, direction-weighted complex network for the flame image;

[0107] Based on a complex network, a derivative information weighting formula reflecting the continuous change of vertices at different scales is defined. By integrating directional information, a multi-scale irregular directional weighted color texture complex network descriptor for furnace flame images is constructed. Combining the color channel information of each color channel in the HSI space of the flame image and the high-order local derivative, a directional weighted multi-layer network mode is constructed to enhance the description of color information, thereby better dealing with the high similarity of random natural textures presented by flame images under different carbon contents at the steelmaking endpoint.

[0108] Example 3

[0109] like Figure 1-3

[0110] Step 3.3: Calculate the directional weighted feature quantization of the complex network topology connection pattern of the flame image to describe the vertex directional connection information, construct the color texture feature of the furnace mouth flame image, and extract multi-scale irregular directional detail information of the furnace mouth flame to obtain more effective texture difference features to adapt to the changes of the flame.

[0111] Step 3.3.1: Define the vertex orientation weighting degree of the complex network for the flame image as the vertex orientation weighting degree. The formula, obtained by summing the direction parameters of the other connected vertices, is as follows:

[0112] (11)

[0113] In the formula, For the direction parameter, this paper sets... To achieve a set of parallel trend vertex direction parameters summing to 1, Used to compute each vertex of the flame image texture network The directional weighting describes the texture changes of the flame.

[0114] By computing each vertex of the texture network degree , histogram degree The calculation is shown in equation (12):

[0115] (12)

[0116] in Defined by equation (13):

[0117] (13)

[0118] Step 3.3.2: Analyzing the degree histogram of the complex network in the flame image can yield numerous features that can be used as multi-directional, multi-scale, and irregular color texture features of the flame image.

[0119] Calculate the directional weighting degree of complex networks The statistical features corresponding to its histogram are used as color texture descriptors for flame images. The feature vectors describing each layer of the network in the flame image are... As shown in equation (14), the flame image DDMCN color texture descriptor is constructed by fusing multi-layer network information.

[0120] (14)

[0121] in, It is a threshold. The calculation yielded the result.

[0122] (15)

[0123] Step 4: Input the extracted color texture features into the KNN regression model to predict the endpoint carbon content, and obtain a relatively accurate and stable converter endpoint carbon content prediction result through five-fold cross-validation.

[0124] The directional weighted feature describing vertex directional connectivity information is calculated to quantify the complex network topology connectivity pattern of flame images. Color texture features of furnace mouth flame images are constructed, and multi-scale irregular directional detail information of the furnace mouth flame is extracted to obtain more effective texture difference features to adapt to flame variations. A KNN regression model is established to predict the endpoint carbon content.

[0125] Experimental results show that the carbon content prediction accuracy of this invention reaches 94.13% within an error range of ±0.02%. The carbon content prediction results are as follows: Figure 3 As shown, this invention effectively solves the problem of high similarity between flame images corresponding to different carbon contents at the converter steelmaking endpoint due to the complexity of flame changes and the difficulty in extracting multi-directional, multi-scale, and irregular color texture features from the furnace mouth flame image. This makes it difficult to distinguish between flame images with similar carbon contents, thus hindering accurate carbon content prediction. The extracted irregular color texture features of the furnace mouth flame have strong discriminative power, providing a reference for research on methods to improve the accuracy of carbon content prediction in molten steel at the converter steelmaking endpoint. It can also provide a reference for image processing, such as flame images, which have high similarity and contain directional information, and involve multi-scale irregular texture extraction.

[0126] The above description, in conjunction with the accompanying drawings, provides a detailed account of specific embodiments of the present invention, which is merely for the purpose of illustrating the invention. The preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific embodiments described. Obviously, based on the content of this specification, various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the invention. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for predicting the final carbon content in converter steelmaking based on feature extraction from flame images using a graphical model. The specific steps are as follows: Step 1: Collect video footage of the flames at the furnace opening in different heats during actual converter steelmaking production, captured by industrial cameras. Step 2: Sample the captured video frame by frame to obtain the flame image dataset, which will be used as experimental data. The label is the carbon content value of molten steel measured by the secondary lance detection technology at the end of the converter steelmaking process. Its features are: Step 3: Use a derivative nonlinear mapping direction-weighted multilayer complex network to extract multi-scale irregular directional detail information of the flame at the furnace mouth, and describe the color texture features of the flame in the flame image dataset; Step 4: Input the extracted color texture features into the KNN regression model to predict the endpoint carbon content, and obtain the prediction results of the converter endpoint carbon content through five-fold cross-validation. Step 3 specifically involves: Step 3.1: After removing the furnace mouth boundary information, construct the flame image phase space mapping map; Step 3.2: Based on complex networks, define a derivative information weighting formula that reflects the continuous changes of vertices at different scales. Integrate directional information to construct a multi-scale, irregularly oriented, weighted color texture complex network descriptor for the furnace flame image. Combine the color channel information and higher-order local derivatives of the flame image's HSI space to construct an oriented weighted multilayer network pattern. Model the phase space map of the flame image into a network. If a flame image exists... Phase space map of a certain color channel conversion width and height common 1 pixel, i.e. For each vertex, the phase value of each pixel is used as the vertex value. Network connection rules are constructed by combining the derivative information between vertices, the weight formula, and the direction information. An undirected weighted network is built according to these connection rules, and combined with color information, thus constructing a multi-layered mapped direction-weighted complex network. ; Step 3.3: Calculate the directional weighted feature quantization of the complex network topology connection pattern of the flame image to describe the vertex directional connection information, construct the color texture feature of the furnace mouth flame image, and extract multi-scale irregular directional detail information of the furnace mouth flame.

2. The method for predicting the carbon content at the end point of converter steelmaking by extracting features from a graph model flame image according to claim 1, characterized in that, In step 3.1, the specific steps for constructing the flame image phase space mapping map after removing the furnace mouth boundary information are as follows: Step 3.1.1: Remove the parts other than the flame by segmenting using the Otsu's method; Step 3.1.2: Construct spatial mutual information of local textures, and further describe the complex spatial changes of flame textures through phase form; calculate phase space maps in H, S, and I color channels respectively to statistically analyze the structural differences of local flame regions; To preserve useful information about local flame characteristics from the original image, the image is divided into several non-overlapping segments. The flame image can be described in phase form by dividing the image into small blocks and calculating the phase value within each block. Its mathematical expression is Equation (1): (1) In the formula, The average value of the pixels surrounding the center point of a local block in a flame image is expressed by equation (2). Indicates the intensity value of surrounding pixels. The transformed phase space map is represented by integrating surrounding pixels and the average value to form the spatial domain representation of the flame image. (2)。 3. The method for predicting the carbon content at the end point of converter steelmaking by extracting features from a graph model flame image according to claim 1, characterized in that, In step 3.2, the specific steps for constructing a multi-scale, irregularly oriented, weighted color texture complex network descriptor for the furnace flame image are as follows: Considering the derivative information between a set of vertices, the vertex pairs of the flame image phase space map complex network are... and , The connection weight between them is defined as and , The intensity difference between them, after being mapped by the derivative, is calculated as shown in equation (3-4): (3) (4) In the formula As vertex and , The first-order derivative relationship and the second-order derivative mapping relationship between them are shown in (5-6), respectively. The connection weights represent the first-order derivative relationships between vertices. The connection weights represent the second derivative relationships between vertices; (5) (6) In the formula This represents the phase value at the center vertex. express Phase values ​​of adjacent scale vertices, This represents the nonlinear mapping value of the first derivative after phase transformation of the flame image. and express Phase values ​​of adjacent scale vertices, This represents the transformed second derivative nonlinear mapping value. n Defined in equation (7), represents the mean; (7) Consider the spatial orientation relationships of a set of vertices, that is, the mutual changes of the flame in different directions; if there exists a Local region, with vertices Taking the eight-neighbor domain as an example, the direction is represented by the sequence number. The formula is expressed as (8): (8) calculate Serial number is Vertex in direction Pairs with neighboring points The difference, after being mapped nonlinearly by the derivative, yields the following difference pair: , for equation (9): (9) Combining information about the direction of flame changes, vertex changes are used as network connection conditions; the direction of the arrow indicates a pixel change from small to large. A parallel trend is satisfied when the center vertex and neighboring vertices change from small to large or from large to small; a non-parallel trend is satisfied when all neighboring vertex values ​​are greater than or less than the center vertex. In other words, when the vertex... and neighboring point pairs When the change satisfies a parallel trend, then... Establish network connections; conversely, when changes satisfy a non-parallel trend, then... No network connection is established, which defines it as The connection method between the vertex and the surrounding vertices is expressed mathematically as equation (10): (10) In the formula, A value of 1 indicates a connection, while 0 indicates no connection; where, ; Each pixel of the phase spectrum of each color channel is considered a vertex, and the total number of vertices is... , forming a set , forming the edge Superimposing three color channels yields a multi-layered, direction-weighted complex network for the flame image.

4. The method for predicting the carbon content at the end point of converter steelmaking by extracting features from a graph model flame image according to claim 1, characterized in that, In step 3.3, the specific steps for constructing the color texture features of the furnace flame image are as follows: Step 3.3.1: Define the vertex orientation weighting degree of the complex network for the flame image as the vertex orientation weighting degree. The summation of the direction parameters of the other connected vertices yields the result, which is expressed by formula (11): (11) In the formula, For the direction parameter, this paper sets... To achieve a set of parallel trend vertex direction parameters summing to 1, Used to compute each vertex of the flame image texture network Directional weighting degree; By computing each vertex of the texture network degree , histogram degree The calculation is shown in equation (12): (12) in Defined by equation (13): (13) Step 3.3.2: Analyzing the degree histogram of the complex network in the flame image can yield numerous features that can be used as multi-directional, multi-scale, irregular color texture features of the flame image. Calculate the directional weighting degree of complex networks The statistical features corresponding to its histogram are used as color texture descriptors for flame images; the feature vectors of each layer of the network describe the flame image. Equation (14); The flame image DDMCN color texture descriptor is constructed by fusing multi-layer network information. (14) in, It is a threshold. Calculations show that (15)。

Images