Automatic rating method for center segregation defect of continuous casting billet

Abstract

The invention discloses an automatic rating method for center segregation defects of a continuous casting billet. According to the scheme, the method comprises the steps that median filtering processing is conducted on an input center segregation macrostructure image of the continuous casting billet; performing image enhancement based on Retinex reflection transformation on the filtered image, adjusting the brightness and contrast of the image, retaining details and eliminating the influence of illumination conditions; carrying out binarization processing on the enhanced image by adopting an improved Otsu threshold algorithm to generate a binarized image; performing cutting and connected domain analysis on the binarized image, and removing boundary interference and noisy points of which the areas are smaller than a set threshold value; extracting geometric features and distribution features of the center segregation defect from the processed image; performing correlation analysis on the extracted features by using an SHAP algorithm, screening key features and removing redundant features; and based on SHAP value weight distribution of the key features, constructing a rating model and outputting a final rating result of the center segregation defect.

Description

Technical Field

[0001] The present invention relates to the technical field of continuous casting slab defect detection, and in particular to an automatic rating method for continuous casting slab center segregation defects. Background Art

[0002] Centerline segregation in continuous casting slabs is a major type of steel quality defect, closely related to continuous casting process parameters. Traditional classification methods rely on manual visual comparison using the Mannesmann classification system, which suffers from significant subjective errors, high sensitivity to illumination, and loss of detail. While existing technologies, such as patent CN201710852006.5, propose quantitative characterization methods, noise interference and uneven illumination in image processing still lead to inaccurate feature extraction.

[0003] Therefore, an automated rating method for the center segregation defect of continuous casting slabs is urgently needed to improve the objectivity and reliability of the rating. Summary of the Invention

[0004] The present invention provides an automatic rating method for central segregation defects of continuous casting billets, which is used to overcome at least one technical problem existing in the prior art.

[0005] An embodiment of the present invention provides an automatic rating method for center segregation defects in continuous casting slabs, comprising:

[0006] S1. Performing median filtering on the input low-magnification image of the center segregation of the continuous casting slab to be analyzed to reduce image noise;

[0007] S2. Perform image enhancement based on Retinex reflection transformation on the image after median filtering to adjust the brightness and contrast of the image. The specific formula is:

[0008] log R(x,y)=log[I(x,y)]-log[L(x,y)]

[0009] Wherein, the symbol I(x,y) represents the original image, the symbol L(x,y) represents the incident light component, and the symbol R(x,y) represents the enhanced reflected component;

[0010] S3, using the improved Otsu threshold algorithm to perform binarization processing on the image after the reflection transformation to generate a binary image;

[0011] S4, performing cropping and connected domain processing on the binary image to eliminate boundary effects and remove noise points based on area;

[0012] S5. Extracting geometric features and distribution features for center segregation defect analysis from the image processed in step S4;

[0013] S6. Use the SHAP value method to perform correlation analysis on the features extracted in step S5, screen key features and remove redundant features;

[0014] S7. Assign weights based on the SHAP values of key features, build a rating model, and output the rating results of center segregation defects.

[0015] In an optional embodiment, before performing the median filtering process in step S1, the step includes: converting the input low-magnification microstructure image of the continuous casting slab center segregation into a grayscale image, and then performing a standardization process. The formula for the standardization process is:

[0016]

[0017] in, The symbol μ represents the mean of the image, the symbol X represents the image matrix, the symbol σ represents the standard deviation, and the symbol N represents the number of pixels in the image;

[0018] Then, median filtering is performed based on the normalized image.

[0019] In an optional embodiment, before performing the image enhancement based on the Retinex reflection transformation on the image processed by the median filter, the method further includes: performing Gamma correction on the image processed by the median filter to obtain a Gamma-corrected image;

[0020] Then, the image after the Gamma correction is subjected to the Retinex reflection transformation.

[0021] In an optional embodiment, the specific content of step S3 includes:

[0022] An initial threshold T0 is set, and the image after the reflection transformation is divided into two parts, the foreground area F and the background area B, according to the threshold.

[0023] The total number of pixels in the image after reflection transformation is N, of which the number of foreground pixels is N f , the number of background pixels is N b ;

[0024] Determine the total gray level of the image after reflection transformation as L-1, and count the number of pixels at each gray level i as N i ;

[0025] Calculate the probability that the total number of foreground and background pixels accounts for the total number of pixels, where the foreground probability Background probability

[0026] Calculate the grayscale average of the foreground and background, the foreground grayscale average Background gray average

[0027] Calculate the grayscale average value M=P of the entire image f ×M f +P b ×M b ;

[0028] Calculate the inter-class variance σ between foreground and background 2 (T)=P f ×(M f -M) 2 +P b ×(M b -M) 2 ;

[0029] Keep adjusting the threshold and repeat the above steps until you find the between-class variance σ 2 (T) The final segmentation threshold T that reaches the maximum value optimal ;

[0030] The final threshold T optimal The image is binarized to obtain the binarized image.

[0031] In an optional embodiment, the geometric features include the total number, area, length, width, maximum length, maximum width, maximum area, average length of the central segregation point, average width of the central segregation point, total segregation area, column continuity of the segregation band, and row continuity of the segregation band of the continuous casting billet;

[0032] The distribution characteristics include the maximum peak, peak area, half-height width, and number of peaks in the full width of the row scan spectrum, and the maximum five peaks, peak positions, and maximum breakpoint area of the column scan spectrum.

[0033] In an optional embodiment, in step S6, the contribution of each feature to the rating result is quantified by the SHAP value, the top N features in importance are selected as key features, and redundant features whose contribution is lower than a set threshold are eliminated.

[0034] In an optional embodiment, the specific content of step S7 includes:

[0035] The SHAP value is normalized using the formula:

[0036]

[0037] Among them, the symbol P i Represents the probability distribution value of the i-th feature, symbol φ i represents the corresponding SHAP value, and the symbol N represents the total number of features;

[0038] The normalized SHAP value is mapped to a probability distribution through the softmax function, and the formula is:

[0039]

[0040] Among them, the symbol ω i represents the rating weight of the i-th feature;

[0041] A weighted scoring model is constructed based on the rating weights to obtain a rating score for the low-magnification microstructure image of the continuous casting slab center segregation to be analyzed; and a rating result of the continuous casting slab center segregation defect is determined based on the rating score and a preset rating threshold. The formula of the weighted scoring model is: Among them, the symbol F k represents the normalized value of the i-th feature.

[0042] In an optional embodiment, determining the rating result of the continuous casting slab center segregation defect based on the rating score and a preset rating threshold specifically includes:

[0043] The defect grade of the central segregation defect of the continuous casting slab is divided according to the preset rating thresholds T1 and T2. If the rating score is ≤T1, it is judged as a mild defect; if T1<rating score≤T2, it is judged as a moderate defect; if the rating score is ≥T2, it is judged as a severe defect. T1 and T2 are determined by historical data.

[0044] In an optional embodiment, the normalization of the SHAP value includes performing a logarithmic transformation on the absolute value of the SHAP value based on the following formula:

[0045] φ' i =log(1+|φ i |);

[0046] Then further normalization is performed based on the following formula:

[0047]

[0048] Among them, φ i is the SHAP value of the original i-th feature, φ' i is the i-th eigenvalue after logarithmic transformation, P i is the final probability distribution value of the i-th feature, and N is the total number of features.

[0049] In an optional embodiment, the method for determining the preset rating thresholds T1 and T2 includes:

[0050] Acquiring historical rating data, the data including rating scores of continuous casting slab center segregation defects and corresponding defect grade categories;

[0051] The K-means algorithm is used to classify the rating scores into three categories. Three cluster center values are obtained through iterative calculation, and two of the center values are used as the initial thresholds T1 and T2;

[0052] Based on ROC curve analysis, the classification accuracy of different threshold combinations is traversed, and T1 and T2 that maximize the accuracy are selected as the final defect level classification thresholds.

[0053] One embodiment of this specification can at least achieve the following beneficial effects: In the technical solution of this application, in terms of image processing, an optimization solution combining the Retinex algorithm and the Otsu threshold algorithm is adopted to realize automatic threshold adjustment, effectively solve the problems of uneven illumination and poor noise removal, retain image details, improve the quality and clarity of image processing, and provide an accurate image basis for subsequent ratings. In terms of feature extraction, this method not only comprehensively extracts multiple features of center segregation defects, but also ensures the accuracy and efficiency of ratings through feature importance, validity proof and redundant feature removal. Finally, the technical solution of this application assigns weights based on the SHAP values of key features, constructs a rating model and outputs the rating results of center segregation defects, so that the technical solution of this application can accurately and automatically evaluate the center segregation defects of continuous casting billets, reduce subjective errors, and provide reliable technical support for the steelmaking continuous casting process. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0055] Figure 1 A flow chart of an automatic rating method for center segregation defects in continuous casting slabs provided by an embodiment of the present invention;

[0056] Figure 2 This is the program running process interface of the automatic rating method for the center segregation defect of the continuous casting slab in an embodiment of the present invention;

[0057] Figure 3 This is an image processing effect diagram provided by an embodiment of the present invention;

[0058] Figure 4 A schematic diagram showing the improvement and comparison of the effect diagram after image processing provided by an embodiment of the present invention;

[0059] Figure 5 This is the first effect diagram of extracting feature correlation analysis in an embodiment of the present invention;

[0060] Figure 6 This is the second effect diagram of extracting feature correlation analysis in an embodiment of the present invention. DETAILED DESCRIPTION

[0061] To make the purpose, technical solutions, and advantages of one or more embodiments of this specification more clear, the technical solutions of one or more embodiments of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of one or more embodiments of this specification.

[0062] It should be understood that although the terms first, second, third, etc. may be used in this application document to describe various information, these information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.

[0063] Before introducing the technical solution of this application, the concept of center segregation defect of continuous casting billet is explained first. In steel production, the billet formed by cooling and solidifying molten steel in a continuous casting machine is a continuous casting billet. Center segregation defect refers to the uneven chemical composition of the center area of the billet during the solidification process of the continuous casting billet due to factors such as the difference in solubility of elements in the liquid phase and the solid phase and solidification shrinkage. Certain elements (such as carbon, sulfur, manganese, etc.) are enriched or depleted. This phenomenon is called center segregation defect. Center segregation defect will change the organizational structure of steel and make the performance of steel uneven. For example, it will reduce the strength, toughness and fatigue performance of steel, affect the processing performance of steel, and cause defects such as cracks in subsequent rolling, forging and other processes. Since center segregation defect has a great impact on steel quality, it is crucial to accurately detect and evaluate this defect.

[0064] The technical solution of the present application provides an automatic rating method for the central segregation defects of continuous casting billets, which is used to solve the problems of strong subjectivity and inaccurate feature extraction in traditional manual rating. The technical solution of the present application is described in detail below with reference to the accompanying drawings.

[0065] The process of the automatic rating method for center segregation defects of continuous casting slabs provided by one embodiment of the present invention may include the following steps.

[0066] Step S1: performing median filtering on the input low-magnification image of the center segregation of the continuous casting billet to be analyzed to reduce image noise.

[0067] In this step, the low-magnification structural image of the center segregation of the continuous casting billet is a cross-sectional or longitudinal cross-sectional image of the continuous casting billet obtained by macroscopic imaging means (such as low-magnification photography with a metallographic microscope, industrial CT scanning, or ultrasonic imaging), which focuses on presenting the compositional segregation characteristics of the central area (such as the morphology, distribution range, and continuity of the segregation band).

[0068] At the same time, in this step, depending on the image content and noise reduction effect, different filters can be used to filter the input low-magnification image of the central segregation of the continuous casting billet to be analyzed, such as median filtering, bilateral filtering, and Gaussian filtering. Depending on the noise, the filtering effect is also different. Mean filtering will weaken the noise part to the surrounding pixels, resulting in a reduction in noise amplitude, but the particle area of the noise point will also increase, so the contaminated area will increase. The Gaussian filter replaces the pixel value of the pixel neighborhood with the weighted mean of the pixel neighborhood, and the weight of each neighborhood pixel increases and decreases monotonically with the distance from the point to the center. This property is very important because the edge is a local feature of the image. If the smoothing operation still has a significant effect on pixels far from the operator center, the smoothing operation will distort the image. Therefore, the disadvantage is that the weight depends entirely on the Euclidean distance between image pixels and has nothing to do with the content of the image. Median filtering is suitable for salt and pepper noise and impulse noise. Because images contaminated by impulse noise and salt-and-pepper noise experience discontinuous grayscale jumps at corresponding locations, the median filter, a nonlinear filtering method, causes less blur than a linear smoothing filter of the same size for these types of random noise and better preserves edges. However, it can cause loss of small objects in the image. Therefore, it is not suitable for images with many points, lines, and spires. The advantage of the bilateral filter is that it preserves edges. Conventional Wiener or Gaussian filters for noise reduction significantly blur edges and are less effective at preserving high-frequency details. As the name suggests, the bilateral filter has an additional Gaussian variance compared to the Gaussian filter. This is based on a spatially distributed Gaussian filter function. Therefore, near edges, pixels farther away do not significantly affect the values of pixels on the edge, thus preserving the values of pixels near the edge. The bilateral filter smoothes the image while preserving edges.

[0069] Step S2: Perform image enhancement based on Retinex reflection transformation on the image after median filtering to adjust the brightness and contrast of the image. The specific formula is:

[0070] log R(x,y)=log[I(x,y)]-log[L(x,y)]

[0071] Wherein, the symbol I(x, y) represents the original image, the symbol L(x, y) represents the incident light component, and the symbol R(x, y) represents the enhanced reflection component.

[0072] Step S3: Using the improved Otsu threshold algorithm to perform binarization processing on the image after the reflection transformation to generate a binary image.

[0073] The binarized image generated in this step can highlight the key information in the image, remove unnecessary grayscale details, make the image features more obvious, and provide a simpler and clearer image basis for subsequent feature extraction, defect detection and other operations.

[0074] Step S4: performing cropping and connected domain processing on the binary image to eliminate boundary effects and remove noise points based on area.

[0075] Step S5: Extracting geometric features and distribution features for center segregation defect analysis from the image processed in step S4.

[0076] In this step, geometric features refer to features related to the shape, size, and other geometric properties of the continuous casting slab's centerline segregation defect, while distribution features describe the distribution pattern of the segregation defect within the image. This solution extracts both geometric and distribution features to describe the centerline segregation defect from multiple dimensions. Combining these features provides a comprehensive and detailed representation of the defect, providing rich information for subsequent analysis.

[0077] Step S6: Use the SHAP value method to perform correlation analysis on the features extracted in step S5, screen key features and remove redundant features.

[0078] In this step, SHAP value, namely Shapley additivity explanation value, originates from the concept of Shapley value in cooperative game theory. In the machine learning model, it calculates the contribution of each feature to the final prediction result by simulating the change in the prediction result when each feature joins or leaves the model. For the analysis of the center segregation defect of continuous casting billets, this method can quantify the influence of each feature extracted from the image (such as the area and shape parameters of the segregation area, etc.) on the rating result, and provide a numerical basis for subsequent feature screening. This step can optimize the feature set for the analysis of the center segregation defect of continuous casting billets by performing feature correlation analysis using the SHAP value method, screening key features and removing redundant features. On the one hand, the retained key features can more accurately reflect the essential characteristics of the center segregation of continuous casting billets and improve the accuracy of the rating; on the other hand, removing redundant features can reduce the computational complexity of the model, reduce training time and resource consumption.

[0079] Step S7: Assign weights based on the SHAP values of the key features, build a rating model, and output the rating results of the center segregation defect.

[0080] In this step, the SHAP (SHapley Additive exPlanations) value is a method used to interpret the prediction results of machine learning models. It measures the contribution of each feature to the model output. When building a rating model based on key features, by calculating the SHAP value of each key feature, the relative importance of each feature in evaluating the center segregation defect of the continuous casting slab can be clarified. The larger the SHAP value, the greater the impact of the feature on the rating result, while the smaller the impact, the smaller the impact. Weights can then be assigned based on the SHAP values of the key features. Specifically, the SHAP values can be normalized and processed to reasonably reflect the importance of each feature in the overall rating. For example, if a key feature has a high SHAP value, it indicates a greater impact on the center segregation defect and is therefore assigned a larger weight; features with lower SHAP values are assigned a relatively smaller weight. In this way, the rating model can more accurately consider the role of each feature. Finally, a rating model can be constructed using weighted key features. For example, machine learning or statistical analysis methods, such as linear regression models, can be used to multiply each key feature by its corresponding weight and then linearly combine them to obtain a comprehensive rating indicator. Alternatively, more complex machine learning algorithms, such as decision trees and neural networks, can be used to input these weighted key features. Through learning and training, the model automatically explores the complex relationships between features, thereby establishing a rating model that can accurately assess the center segregation defect of continuous casting billets. After the rating model is constructed, the relevant data of the continuous casting billet to be rated can be input into the model. The model will calculate and analyze the input data based on the previously determined weights and learned feature relationships, and ultimately output the rating result of the continuous casting billet center segregation defect. This rating result can be a specific numerical value representing the severity of the segregation defect, or a grade classification such as mild, moderate, or severe, thereby providing intuitive continuous casting billet quality information.

[0081] based on Figure 1 The present specification also provides some specific implementation plans of the method, which are described below.

[0082] In an optional embodiment, before performing the median filtering process in step S1, the following steps may be included: converting the input low-magnification microstructure image of the continuous casting slab center segregation into a grayscale image, and then performing a standardization process. The formula for the standardization process is:

[0083]

[0084] in, The symbol μ represents the mean of the image, the symbol X represents the image matrix, the symbol σ represents the standard deviation, and the symbol N represents the number of pixels in the image;

[0085] Then, median filtering is performed based on the normalized image.

[0086] In the technical solution of the present application, median filtering is performed based on the standardized image, which can more effectively remove noise. The standardization process optimizes the data distribution of the image, so that the median filtering algorithm can better identify and process noise points, while retaining important details of the image and improving the clarity and quality of the image.

[0087] In an optional embodiment and technical solution, before performing the image enhancement based on the Retinex reflection transformation on the image processed by the median filter, the method may further include: performing Gamma correction on the image processed by the median filter to obtain a Gamma-corrected image;

[0088] Then, the image after the Gamma correction is subjected to the Retinex reflection transformation.

[0089] In the technical solution of this embodiment, the image after median filtering is first subjected to Gamma correction. Gamma correction is a technique for adjusting the grayscale value of an image through mathematical transformation. Its formula is: out (x,y)=I in (x,y) γ , where symbol I out (x,y) refers to the pixel value of the output image at the (x,y) position, symbol I in (x,y) refers to the pixel value at the (x,y) position in the input image. γ is the gamma value. Calculations first map the pixel value from the typical 0-255 range to the 0-1 range, and then map it back to the 0-255 range. The gamma value also affects the image quality: a value less than 1 enhances low-brightness areas, while a value greater than 1 enhances high-brightness areas.

[0090] In the technical solution of this application, gamma correction is used as a preprocessing step to prioritize adjusting the overall brightness distribution of the image, providing a more balanced input for the subsequent Retinex algorithm. Gamma correction first globally adjusts the brightness (such as improving the contrast in dark areas) to avoid noise amplification caused by inaccurate illumination estimation in extremely dark areas by Retinex. The gamma-corrected image is then subjected to a Retinex-based reflection transformation. Retinex post-processing can further eliminate local illumination unevenness and enhance details (such as the morphology and continuity of segregation points).

[0091] In an optional embodiment, the specific contents of step S3 may include:

[0092] An initial threshold T0 is set, and the image after the reflection transformation is divided into two parts, the foreground area F and the background area B, according to the threshold.

[0093] The total number of pixels in the image after reflection transformation is N, of which the number of foreground pixels is N f , the number of background pixels is N b ;

[0094] Determine the total gray level of the image after reflection transformation as L-1, and count the number of pixels at each gray level i as N i ;

[0095] Calculate the probability that the total number of foreground and background pixels accounts for the total number of pixels, where the foreground probability Background probability

[0096] Calculate the grayscale average of the foreground and background, the foreground grayscale average Background gray average

[0097] Calculate the grayscale average value M=P of the entire image f ×M f +P b ×M b ;

[0098] Calculate the inter-class variance σ between foreground and background 2 (T)=P f ×(M f -M) 2 +P b ×(M b -M) 2 ;

[0099] Keep adjusting the threshold and repeat the above steps until you find the between-class variance σ 2 (T) The final segmentation threshold T that reaches the maximum value optimal ;

[0100] The final threshold T optimal The image is binarized to obtain the binarized image.

[0101] In an optional embodiment, the geometric features may include the total number, area, length, width, maximum length, maximum width, maximum area, average length and average width of the central segregation points of the continuous casting slab, total segregation area, column continuity of the segregation bands, and row continuity of the segregation bands. The distribution features may include the maximum peak value, peak area, half-height width, and number of peaks within the full width of the row scan pattern, as well as the maximum five peaks, peak positions, and maximum breakpoint area of the column scan pattern.

[0102] In this embodiment, from the perspective of geometric characteristics, the total number of segregation points in the center of the continuous casting billet can intuitively reflect the degree of discreteness of the segregation phenomenon. The more the number, the more dispersed the segregation points. Parameters such as area, length, width, and maximum length, maximum width, and maximum area can describe the size and shape characteristics of individual segregation points. The maximum length and maximum width can reflect the maximum size of the segregation point in different directions and can show the degree of extension of the segregation point; the maximum area highlights the individual scale of the largest area among the segregation points. The average length and average width of the central segregation point are the statistical averages of the lengths and widths of all segregation points, which can reflect the general size of the segregation points and help to understand the overall size trend of the segregation points. The total segregation area is the sum of the areas of all segregation points. It measures the overall size of the segregation area. The larger the total area, the wider the range of the segregation phenomenon in the center area of the continuous casting billet. The column continuity and row continuity of the segregation band are used to evaluate the continuity of the segregation band in the column direction and the row direction. A high degree of continuity means that the segregation band is relatively coherent in the corresponding direction, and a larger segregation area may be formed; a low degree of continuity means that the segregation band is relatively discrete and the continuity of the segregation area is poor.

[0103] From a distribution perspective, the maximum peak value of the row scan spectrum represents the maximum signal intensity of the segregation region during the row scan. The higher the peak value, the more pronounced the segregation region on that row, potentially indicating severe segregation. The peak area reflects the extent of the segregation region along the row direction; a larger area indicates a wider area affected by segregation. The full-width-half-maximum (FWHM) refers to the width at half the height of the peak and measures the concentration of the segregation region. The narrower the FWHM, the more concentrated the segregation region. The number of peaks within the full-width spectrum reflects the number of peaks across the entire width of the row scan spectrum. A greater number of peaks indicates a more frequent distribution of segregation along the row direction. The five largest peaks and their positions in the column scan spectrum focus on the regions with the highest segregation signal intensity along the column direction and their locations. The five largest peaks highlight regions with severe segregation along the column direction, while the peak positions clarify the specific location of these regions along the column direction. The maximum breakpoint area indicates the largest area where the segregation distribution appears discontinuous in the column scan spectrum. It reflects the continuity of segregation in the column direction. The larger the breakpoint area, the more obvious the discontinuity of segregation in the column direction and the worse the continuity of segregation distribution.

[0104] By selecting the geometric characteristics and distribution characteristics mentioned above, the morphology and severity of segregation defects can be fully described.

[0105] In an optional embodiment, the contribution of each feature to the rating result is quantified by the SHAP value in step S6, the top N features in importance are selected as key features, and redundant features whose contribution is lower than a set threshold are eliminated.

[0106] In the technical solution of this embodiment, the SHAP value is a method for interpreting the prediction results of a machine learning model. It can measure the contribution of each feature to the model output results. In the continuous casting slab center segregation defect rating model, each feature extracted from the image, such as the area of the segregation point, the maximum peak value of the line scan spectrum, etc., has different degrees of influence on the final rating result. The SHAP value quantifies the influence of each feature in a specific numerical value through complex calculations.

[0107] After quantifying the contribution of all features, the technical solution of this embodiment ranks these features according to their importance, and selects the top N features in importance from a large number of features as key features. The value N here can be set according to actual conditions and experience. These key features contain the most representative information that has the greatest impact on the rating results. In the technical solution of this embodiment, after screening the key features, redundant features whose contribution is lower than the set threshold are also eliminated. The threshold is set to distinguish which features have a negligible impact on the rating results. These redundant features may be highly correlated with other features, or have an extremely weak impact on the rating results. For example, the SHAP value of a feature is very small, far below the set threshold, indicating that it provides very little effective information in the rating process. Retaining it will not only increase the computational complexity of the model, but may also introduce noise, affecting the accuracy of the rating. Therefore, eliminating such features can simplify the model, improve the model's operating efficiency and the accuracy of the rating.

[0108] In an optional embodiment, the specific content of step S7 may include:

[0109] The SHAP value is normalized using the formula:

[0110]

[0111] Among them, the symbol P i Represents the probability distribution value of the i-th feature, symbol φ i represents the corresponding SHAP value, and the symbol N represents the total number of features;

[0112] The normalized SHAP value is mapped to a probability distribution through the softmax function, and the formula is:

[0113]

[0114] Among them, the symbol ω i represents the rating weight of the i-th feature;

[0115] A weighted scoring model is constructed based on the rating weights to obtain a rating score for the low-magnification microstructure image of the continuous casting slab center segregation to be analyzed; and a rating result of the continuous casting slab center segregation defect is determined based on the rating score and a preset rating threshold. The formula of the weighted scoring model is: Among them, the symbol F k represents the normalized value of the i-th feature.

[0116] In an optional embodiment and technical solution, determining the rating result of the continuous casting slab center segregation defect based on the rating score and a preset rating threshold may specifically include:

[0117] The defect grade of the central segregation defect of the continuous casting slab is divided according to the preset rating thresholds T1 and T2. If the rating score is ≤T1, it is judged as a mild defect; if T1<rating score≤T2, it is judged as a moderate defect; if the rating score is ≥T2, it is judged as a severe defect. T1 and T2 are determined by historical data.

[0118] In the technical solution of this embodiment, the thresholds T1 and T2 can be determined by historical data, thereby ensuring the scientific nature and objectivity of the rating standards. This evaluation method based on actual data can avoid subjective arbitrariness and enable the rating results to truly reflect the severity of the center segregation defect of the continuous casting billet.

[0119] In an optional embodiment, the normalization of the SHAP value may include performing a logarithmic transformation on the absolute value of the SHAP value based on the following formula:

[0120] φ' i =log(1+|φ i |);

[0121] Then further normalization is performed based on the following formula:

[0122]

[0123] Among them, φ i is the SHAP value of the original i-th feature, φ' i is the i-th eigenvalue after logarithmic transformation, P i is the final probability distribution value of the i-th feature, and N is the total number of features.

[0124] In the technical solution of this embodiment, the absolute value of the SHAP value is logarithmically transformed, which can compress the wider numerical range of the original SHAP value. When the SHAP value is large, the growth rate of the result after logarithmic transformation slows down, which can prevent individual excessively large SHAP values from dominating the subsequent processing, so that SHAP values of different magnitudes can be compared and processed on a relatively balanced scale.

[0125] Further normalization processing can convert the SHAP value after logarithmic transformation into a probability distribution value, so that the importance of each feature is under the same measurement standard, which facilitates a clear and intuitive comparison of the contribution of different features to the rating results.

[0126] In an optional embodiment, the method for determining the preset rating thresholds T1 and T2 may include:

[0127] Acquiring historical rating data, the data including rating scores of continuous casting slab center segregation defects and corresponding defect grade categories;

[0128] The K-means algorithm is used to classify the rating scores into three categories. Three cluster center values are obtained through iterative calculation, and two of the center values are used as the initial thresholds T1 and T2;

[0129] Based on ROC curve analysis, the classification accuracy of different threshold combinations is traversed, and T1 and T2 that maximize the accuracy are selected as the final defect level classification thresholds.

[0130] In the technical solution of the present application, by obtaining historical rating data containing rating scores and corresponding defect grade categories, a true and reliable basis can be provided for determining the rating threshold, thereby avoiding subjective and arbitrary setting of thresholds and making the thresholds more in line with the actual center segregation of continuous casting billets. At the same time, by using the K-means algorithm to classify the rating scores into three categories and using the obtained cluster center value as the initial threshold, the rating scores can be quickly preliminarily divided, and representative dividing points can be extracted from a large amount of data. Finally, based on the ROC curve analysis, the classification accuracy of different threshold combinations is traversed, and T1 and T2 that maximize the accuracy are selected as the final thresholds. This can fully consider the classification effect under different threshold divisions, maximize the accuracy of defect grade division, and ensure that the rating results can more accurately reflect the true situation of the center segregation defect of the continuous casting billet.

[0131] In the technical solution of the present application, in terms of image processing, an optimization solution combining the Retinex algorithm and the Otsu threshold algorithm is adopted to realize automatic threshold adjustment, effectively solve the problems of uneven illumination and difficult noise removal, retain image details, improve the quality and clarity of image processing, and provide an accurate image basis for subsequent ratings. In terms of feature extraction, this method not only comprehensively extracts multiple features of center segregation defects, but also ensures the accuracy and efficiency of ratings through feature importance, validity proof and redundant feature removal. Finally, the technical solution of the present application assigns weights based on the SHAP values of key features, constructs a rating model and outputs the rating results of center segregation defects, so that the technical solution of the present application can accurately and automatically evaluate the center segregation defects of continuous casting billets, reduce subjective errors, and provide reliable technical support for the steelmaking continuous casting process.

[0132] The following describes the central segregation defect rating program developed based on the technical solution of the previous embodiment in combination with actual cases to illustrate the technical solution of this application, and compares the effects of the technical solution of this application with those of the prior art in combination with the technical effects of the actual cases.

[0133] Example 1

[0134] The program operation interface of the present invention is as follows Figure 2 As shown, Figure 2 The following is the detailed process for developing the standard for the central segregation defect rating method:

[0135] 1. Obtain a low-magnification image of the thickness section of the continuous casting billet center segregation, such as Figure 3 Read in the picture as shown;

[0136] 2. Enter the actual size of the continuous casting billet image in the actual size box (the actual size is the actual length corresponding to the center segregation image of the continuous casting billet);

[0137] 3. Manually enter the noise reduction threshold in the noise reduction threshold box. It is currently set to 30 and can be adjusted later as needed.

[0138] 4. In the sample number box, manually enter the appropriate sample number as required.

[0139] 5. The binarization threshold is automatically set by the adjusted OTSU algorithm and does not need to be entered manually;

[0140] 6. Click the data extraction button to get the processed image (such as Figure 3 The processed image is shown in the figure) and the corresponding feature extraction results (the extracted features are shown in the figure). Figure 1 The feature extraction results can be displayed on the interface as required.

[0141] 7. Click the Save Results button to save all feature extraction results to Excel for subsequent correlation analysis, formulation of rating results, and quantitative characterization of central segregation.

[0142] 8. After saving a large amount of image feature data, use the features as training parameters and the actual rating results as prediction parameters. Use machine learning methods to train the data and input it into the SHAP model to obtain the relative importance results of all features, such as Figure 5 shown.

[0143] For the 20 extracted features, the technical solution of this application uses SHAP value to measure the influence of each feature on the rating result, wherein the algorithm is arranged from top to bottom according to the importance value. Among them, there is a significant correlation between the features such as the maximum peak value of the row scan, the continuity of the segregation band column, the continuity of the segregation band row, and the average width of the segregation point and the final rating result. Among them, the maximum peak value of the row scan, the continuity of the segregation band column, and the average width of the segregation point are positively correlated with the final rating result, while the continuity of the segregation band row is negatively correlated with the final rating result. Several features with little correlation can be further removed to continue the correlation analysis, such as Figure 6 As shown in the figure, the features with the greatest correlation are finally retained for the formulation of a new rating standard, thereby improving the accuracy and reliability of the rating procedure and further optimizing the evaluation process of center segregation defects.

[0144] Example 2

[0145] The obtained low-magnification images of the continuous casting slab thickness cross section were analyzed using the technology disclosed in patent CN201710852006.5 and the present invention respectively;

[0146] 1. Obtain a low-magnification image of the thickness section of the continuous casting billet center segregation, such as Figure 4 As shown in a, b, g, h, and m;

[0147] 2. Use the technology disclosed in patent CN201710852006.5 to process the continuous casting billet image. The results are as follows: Figure 4 As shown in c, d, i, j, n;

[0148] 3. The low-magnification image of the thickness cross section of the continuous casting slab center segregation was processed by the present invention. The extraction steps were the same as 1-6 in Example 1. The processing effect was as follows: Figure 4 As shown in e,f,k,l,o;

[0149] Compared with the method disclosed in patent CN201710852006.5, the present invention solves some new problems. For example, the extraction method disclosed in patent CN201710852006.5 is difficult to meet the feature extraction requirements under different lighting conditions. Specifically, Figure 4As shown in (c), a large number of redundant noise points can be observed, which indicates that the extracted features may contain a large amount of irrelevant information, resulting in feature redundancy. Figure 4 In (d), we further noticed the problem of excessive noise and inaccurate feature extraction, especially the redundant noise on the right side, which may have a negative impact on subsequent data processing and result judgment. In addition, Figure 4 (i) It clearly shows that the extraction method may lead to the loss of details, and this loss of details may directly affect the subsequent data analysis and rating judgment. More obviously, Figure 4 The feature loss phenomenon shown in (j,n) further highlights the severity of the problem, as this loss can lead to inaccurate and reduced reliability of subsequent judgment results, thus affecting the effectiveness and credibility of the entire feature extraction. However, as can be seen from the comparison results, the present invention solves the problem of inaccurate feature extraction under different lighting conditions.

[0150] The present invention can better adapt to the feature extraction requirements under different lighting conditions. Figure 3 and Figure 4 As shown in the figure, it can effectively filter the noise in the image and reduce the interference of noise on feature extraction. In the feature extraction process, it can better retain the detailed information in the image and avoid the loss of details. It can better meet the needs of feature extraction and improve the accuracy and reliability of subsequent rating results.

[0151] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic rating method for center segregation defects of continuous casting slabs, characterized in that: The method comprises: S1. Performing median filtering on the input low-magnification image of the center segregation of the continuous casting slab to be analyzed to reduce image noise; S2. Perform image enhancement based on Retinex reflection transformation on the image after median filtering to adjust the brightness and contrast of the image. The specific formula is: logR(x,y)=log[[I(x,y)]]-log[[L(x,y)]] Wherein, the symbol I(x,y) represents the original image, the symbol L(x,y) represents the incident light component, and the symbol R(x,y) represents the enhanced reflected component; S3, using the improved Otsu threshold algorithm to perform binarization processing on the image after the reflection transformation to generate a binary image; S4, performing cropping and connected domain processing on the binary image to eliminate boundary effects and remove noise points based on area; S5. Extracting geometric features and distribution features for center segregation defect analysis from the image processed in step S4; S6. Use the SHAPvalue method to perform correlation analysis on the features extracted in step S5, screen key features and remove redundant features; S7. Assign weights based on the SHAP values of key features, build a rating model, and output the rating results of center segregation defects.

2. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: Before the median filtering process is performed in step S1, the following steps are included: converting the input low-magnification microstructure image of the center segregation of the continuous casting billet into a grayscale image, and then performing a standardization process. The formula for the standardization process is: in, The symbol m represents the mean of the image, the symbol X represents the image matrix, the symbol s represents the standard deviation, and the symbol N represents the number of pixels in the image; Then, median filtering is performed based on the normalized image.

3. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: Before performing the image enhancement based on the Retinex reflection transformation on the image processed by the median filter, the method further includes: performing Gamma correction on the image processed by the median filter to obtain a Gamma-corrected image; Then, the image after the Gamma correction is subjected to the Retinex reflection transformation.

4. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: The specific contents of step S3 include: An initial threshold T0 is set, and the image after the reflection transformation is divided into two parts, the foreground area F and the background area B, according to the threshold. The total number of pixels in the image after reflection transformation is N, of which the number of foreground pixels is N f , the number of background pixels is N b ; Determine the total gray level of the image after reflection transformation as L-1, and count the number of pixels at each gray level i as N i ; Calculate the probability that the total number of foreground and background pixels accounts for the total number of pixels, where the foreground probability Background probability Calculate the grayscale average of the foreground and background, the foreground grayscale average Background gray average Calculate the grayscale average value M=P of the entire image f 'M f +P b 'M b ; Calculate the inter-class variance s between foreground and background 2 (T)=P f ′(M f -M) 2 +P b ′(M b -M) 2 ; Keep adjusting the threshold and repeat the above steps until you find the between-class variance s 2 (T) The final segmentation threshold T that reaches the maximum value optimal ; The final threshold T optimal The image is binarized to obtain the binarized image.

5. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: The geometric characteristics include the total number, area, length, width, maximum length, maximum width, maximum area, average length and average width of the central segregation points, total segregation area, column continuity and row continuity of the segregation bands in the continuous casting billet; The distribution characteristics include the maximum peak, peak area, half-height width, and number of peaks in the full width of the row scan spectrum, and the maximum five peaks, peak positions, and maximum breakpoint area of the column scan spectrum.

6. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: In step S6, the contribution of each feature to the rating result is quantified by the SHAP value, the top N features in importance are selected as key features, and redundant features whose contribution is lower than the set threshold are eliminated.

7. The automatic rating method for center segregation defects of continuous casting slabs according to claim 1, characterized in that: The specific contents of step S7 include: The SHAP value is normalized using the formula: Among them, the symbol P i Represents the probability distribution value of the i-th feature, symbol φ i represents the corresponding SHAP value, and the symbol N represents the total number of features; The normalized SHAP value is mapped to a probability distribution through the softmax function, and the formula is: Among them, the symbol w i represents the rating weight of the i-th feature; A weighted scoring model is constructed based on the rating weights to obtain a rating score for the low-magnification microstructure image of the continuous casting slab center segregation to be analyzed; and a rating result of the continuous casting slab center segregation defect is determined based on the rating score and a preset rating threshold. The formula of the weighted scoring model is: Among them, the symbol F k represents the normalized value of the i-th feature.

8. The automatic rating method for center segregation defects of continuous casting slabs according to claim 7, characterized in that: The step of determining the rating result of the continuous casting slab center segregation defect based on the rating score and a preset rating threshold specifically includes: The defect grade of the central segregation defect of the continuous casting billet is divided according to the preset rating thresholds T1 and T2. If the rating score is ￡T1, it is judged as a mild defect; if T1<rating score ￡T2, it is judged as a moderate defect; if the rating score is ≥T2, it is judged as a severe defect; T1 and T2 are determined by historical data.

9. The automatic rating method for center segregation defects of continuous casting slabs according to claim 7, characterized in that: Before normalizing the SHAP value, the method includes performing a logarithmic transformation on the absolute value of the SHAP value based on the following formula: f′ i =log(1+|φ i |); Then further normalization is performed based on the following formula: Among them, φ i is the SHAP value of the original i-th feature, φ′ i is the i-th eigenvalue after logarithmic transformation, P i is the final probability distribution value of the i-th feature, and N is the total number of features.

10. The automatic rating method for center segregation defects of continuous casting slabs according to claim 8, characterized in that: The method for determining the preset rating thresholds T1 and T2 includes: Acquiring historical rating data, the data including rating scores of continuous casting slab center segregation defects and corresponding defect grade categories; The K-means algorithm is used to classify the rating scores into three categories. Three cluster center values are obtained through iterative calculation, and two of the center values are used as the initial thresholds T1 and T2; Based on ROC curve analysis, the classification accuracy of different threshold combinations is traversed, and T1 and T2 that maximize the accuracy are selected as the final defect level classification thresholds.

Images