Continuous casting billet on-line sizing and weight fixing system based on vision

Abstract

The invention relates to the technical field of ferrous metallurgy continuous casting, and discloses a vision-based continuous casting billet on-line sizing and weight fixing system, which comprises an image acquisition module for acquiring an end face image of a casting billet; the end face detection module is used for processing the end face image, extracting a contour, detecting the defects of squaring and bulging, and calculating the actual end face area; the data synchronization module is used for synchronizing various factor values of the casting blank, including the casting blank steel grade, the pulling speed, the element composition and the outbound temperature; the matching weighing module is used for matching the weighing information with the corresponding casting blank data information; the cut-to-length calculation module is used for calculating the cut-to-length according to the thermal shrinkage rate; the fixed-weight cutting length calculation module is used for calculating and dynamically correcting the cutting length according to the target weight, the end face area and the density; and the display module is used for displaying detection and calculation results in real time. The weight setting precision can be remarkably improved, the weight setting qualification rate reaches 90% or above, full-automatic operation is achieved, and manual intervention is reduced.

Description

Technical Field

[0001] This invention relates to the technical field of continuous casting in iron and steel metallurgy, and more specifically, to a vision-based online length and weight determination system for continuously cast billets. Background Technology

[0002] The length control and weight control of continuously cast billets are critical steps in the steel production process, directly impacting the yield, material utilization, and production costs of subsequent rolling processes. Traditionally, length control typically uses encoders to record the stroke of the straightening machine or laser ranging, combined with preset cutting lengths for billet segmentation. However, due to the presence of different steel grades in the billets and the potential for shrinkage during solidification, relying solely on length measurement can easily lead to deviations in actual weight. Traditional weight control often involves manually pre-setting the length of the billet for cutting. After weighing the cut billet using weighing rollers or offline crane scales, the operator adjusts the cutting length based on weight feedback. However, factors such as vibration and scale can significantly affect weighing accuracy, and changes in steel composition, temperature, and casting speed also influence weight control, with no real-time feedback adjustment possible.

[0003] In recent years, with the development of intelligent manufacturing technology, continuous casting billet length and weight determination has gradually evolved towards multi-sensor fusion and intelligent decision-making. However, existing technologies still face the following challenges: Weighing error: Even if the steel grade, drawing speed and other factors do not change, the same cutting length will have a large weighing fluctuation due to factors such as internal shrinkage cavities.

[0004] Weighing feedback speed: The fixed weight signal needs to be fed back to the cutting equipment in a very short time. However, in existing systems, especially in the case of steelmaking without weighing rollers and other online weighing, the feedback of the weighing weight is often delayed, which is not conducive to manual real-time adjustment of the cutting length.

[0005] Manual operation: Manually adjusting the cutting length of continuously cast billets is affected by subjective factors such as operator experience and fatigue. It requires frequent machine stops for measurement or manual parameter input, making it difficult to adapt to continuous production rhythms and reducing overall production efficiency. Differences in operating habits among different shifts can lead to inconsistent cutting lengths and increased weight fluctuations in the same batch of billets, which is not conducive to standardized production.

[0006] Cross-sectional area measurement: The end face area of ​​continuously cast billets varies with different steel grades, casting speeds, and temperatures. It is difficult to accurately detect the cross-sectional dimensions of billets that are shaped, bulging, or irregularly shaped in real time. Summary of the Invention

[0007] To address the aforementioned technical problems in related technologies, this invention provides a vision-based online length and weight determination system for continuously cast billets, which can solve the above problems.

[0008] To achieve the above-mentioned technical objectives, the technical solution of the present invention is implemented as follows: A vision-based online length and weight determination system for continuously cast billets includes: The image acquisition module is used to acquire images of the end face of the cast billet; The end face detection module is used to process the end face image, extract the contour and detect defects such as squareness and bulging, and calculate the actual end face area. The data synchronization module is used to synchronize the values ​​of various factors of the billet, including the billet steel grade, casting speed, elemental composition, and outlet temperature. The matching weighing module is used to match the weighing information with the corresponding billet data information; The fixed-length cutting calculation module is used to calculate the fixed-length cutting based on the heat shrinkage rate; The constant weight cutting length calculation module is used to calculate and dynamically correct the cutting length based on the target weight, end face area, and density. The display module is used to display the detection and calculation results in real time.

[0009] Furthermore, the image acquisition module includes an end face camera. When the laser temperature sensor detects that the billet to be cut has reached the designated position, the PLC sends the arrival signal to the vision server, and the vision server activates the end face camera to take a picture.

[0010] Furthermore, the end face detection module includes an image preprocessing unit, a semantic segmentation unit, a contour extraction unit, and an area calculation unit. When the billet travels to the designated position in the image, end face detection is performed by the end face detection module. The image preprocessing unit preprocesses the brightness of the billet end face image, converting it into an image with uniform brightness. The semantic segmentation unit calculates the image region where the billet end face image is located by the brightness-adjusted image, and the contour extraction unit extracts the edge contour of the image region. The area calculation unit performs perspective transformation on the image region through planar calibration, so that each pixel represents the same area after the transformation. The number of pixels in the transformed image region is calculated to calculate the actual area value of the billet end face.

[0011] Furthermore, after extracting the edge contour of the image area, contour smoothing is used to remove contour burrs, the four corner points of the contour and the contour center line are found, the difference between the two diagonals of the contour is used to determine whether the end face of the billet is distorted, and the difference between the center line of the contour and the center line of the set center line threshold is used to determine whether the end face of the billet is bulging.

[0012] Furthermore, the billet steel grade refers to the type of billet, the casting speed is the speed at which the billet exits the crystallizer, the elemental composition is the elemental composition of the molten steel after refining, and the exit temperature is the furnace temperature when the molten steel exits the tundish.

[0013] Furthermore, after the billet is cut and enters the conveyor roller, a billet code and flow number will be generated. After the billet is weighed, the same code and corresponding flow number will be generated, matching the billet steel grade, casting speed, element composition, exit temperature and billet weighing weight.

[0014] Furthermore, the thermal shrinkage rate R of the billet and the fixed-length cutting length L are predicted based on the billet steel grade, casting speed, elemental composition, and outlet temperature. 定尺切割 =L 定尺 *S, where L 定尺 This indicates the length value that the cast billet must reach after cooling, as specified by the production plan or order.

[0015] Furthermore, the fixed-weight cutting length calculation module performs the calculation through the following steps: When the billet weight is not received, based on the billet steel grade, casting speed, and elemental composition, the density ρ of the billet is calculated using the random forest algorithm, and the fixed weight W of the billet is determined on-site. 定重 The initial cutting length L1 of the billet is calculated based on the actual end face area A1 measured in real time using vision, and L1 = W. 定重 / (A1*ρ); After receiving the weight of the previously cut billet, the cutting length of the uncut billet is fine-tuned. The fine-tuned cutting length is L2, L2=L1+(W 定重 -W 称重 ) / (A1*ρ)。

[0016] Furthermore, if the drawing speed, end face area, or elemental composition of the blank to be cut and the weighed blank change (and the change exceeds the set change threshold), the cutting length is readjusted based on the drawing speed change coefficient and the elemental composition change coefficient. The readjusted cutting length is L3. Where A2 represents the end face area of ​​the billet to be cut, A3 represents the end face area of ​​the weighed billet used as a reference, ΔB1, ΔB2, ΔB3 represent the changes in casting speed, sulfur content, and phosphorus content, respectively, and K1, K2, K3 represent the coefficients of change in casting speed, sulfur content, and phosphorus content, respectively.

[0017] Furthermore, the system also includes an alarm module for triggering an alarm when severe end face deformation or abnormal weight is detected.

[0018] The beneficial effects of this invention are: This application can significantly improve the accuracy of fixed weight, with a fixed weight qualification rate of over 90%, realize fully automated operation, reduce manual intervention, have intelligent anomaly handling capabilities, and reduce material waste and labor costs. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in 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.

[0020] The present invention will now be described in further detail with reference to the accompanying drawings.

[0021] Figure 1 This is a schematic diagram of a vision-based online length and weight determination system for continuous casting billets, as described in an embodiment of the present invention. Figure 2 This is a normal image of the end face of a cast billet detected by the end face detection module described in this embodiment of the invention; Figure 3 This is an image of the de-squared casting end face detected by the end face detection module described in this embodiment of the invention; Figure 4 This is an image of the bulging end face of the cast billet detected by the end face detection module described in this embodiment of the invention. Detailed Implementation

[0022] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention are within the scope of protection of the present invention.

[0023] like Figure 1As shown, this invention discloses a vision-based online length and weight determination system for continuously cast billets, comprising: an image acquisition module for acquiring images of the billet end face; an end face detection module for processing the end face images, extracting contours, detecting defects such as squareness and bulging, and calculating the actual end face area; a data synchronization module for synchronizing the values ​​of various factors of the billet, including the billet steel grade, casting speed, elemental composition, and outlet temperature; a matching weighing module for matching the weighing information with the corresponding billet data information; a length determination cutting length calculation module for calculating the length determination cutting length based on the thermal shrinkage rate; a weight determination cutting length calculation module for calculating and dynamically correcting the cutting length based on the target weight, end face area, and density; and a display module for displaying the detection and calculation results in real time. This application system switches between fixed-length cutting and fixed-weight cutting based on the on-site cold feeding of cast billets to the cooling bed (process: high-temperature cast billets cut from the continuous casting machine → conveyed to the cooling bed via conveyor rollers → transferred to the cooling bed by the steel transfer machine → naturally or forcibly cooled to room temperature on the cooling bed → off the production line, inspection, stacking, and warehousing) and hot feeding rolling (process: high-temperature cast billets cut from the continuous casting machine → conveyed directly to the steel rolling furnace via insulated rollers or special insulated cars → slightly homogenized or supplemented in the heating furnace → rolled).

[0024] Example 1: This application installs a laser temperature sensor at a designated location on the billet production site and connects it to a PLC. An end-face camera (industrial area array camera) is connected to a vision workstation. The PLC receives the signal from the temperature sensor in real time via TCP communication. When the temperature is detected to be greater than the set temperature threshold, it continuously sends the billet arrival signal to the vision server, and the vision server activates the end-face camera to take pictures.

[0025] Example 2: When the end-face camera receives the start signal, it begins to take continuous pictures and sequentially transmits the photos to the end-face detection module. The end-face detection module detects the position of the billet end face using relevant recognition algorithms until the billet end face reaches the designated position in the image (i.e., the center of the image). Due to inconsistent billet temperatures, the brightness of the images varies. First, the image preprocessing unit preprocesses the brightness of the billet end face image, converting it into an image with uniform brightness. The semantic segmentation unit then calculates the image region containing the billet end face image, and the contour extraction unit extracts the edge contour of the image region. Contour smoothing removes burrs (including inner and outer curved surface burrs), finding the four corner points and the center line of the contour. The lengths of the two diagonals (d1 and d2) are obtained from the four corner points. The difference between the two diagonals, Δd = d1 - d2, is considered a deviation if the difference exceeds a predetermined threshold (e.g., ...). Figure 3 As shown), the difference in the midline is ΔL = L. mid -L0, if the centerline of the contour (L midIf the difference between the centerline and the set centerline threshold (L0) is greater than a specified threshold, it is considered a bulge (e.g., ...). Figure 4 As shown), if Δd ≤ the specified threshold and ΔL ≤ the specified threshold, then it is normal (e.g. Figure 2 As shown in the figure, the area calculation unit performs perspective transformation on the image area through planar calibration, so that each pixel represents the same area after the transformation, calculates the number of pixels in the transformed image area, and calculates the actual area value of the billet end face.

[0026] The image preprocessing unit processes images through the following steps: Image acquisition: Raw RGB image of the end face of the high-temperature billet captured by an industrial area array camera (due to the different cooling rates of the billet edge, corner and center, the image usually shows an uneven brightness field with the brightest central area and darker edges and corners).

[0027] Color space conversion (RGB->HSV / HSI / Lab): In the RGB color space, brightness (lightness) and color information (hue, saturation) are coupled together. In order to adjust brightness independently, it is necessary to convert to a color space that separates brightness (Value / Lightness / Intensity). Common choices are HSV, HSI, and Lab.

[0028] Luminance channel (V or L channel) extraction and modeling: Separate the luminance component map from the converted image (e.g., the V channel image in HSV space, or the L channel image in Lab space). Apply a large-radius Gaussian low-pass filter or morphological opening operation (erosion followed by dilation, using a large structuring element) to the luminance channel map to smooth out high-frequency luminance variations caused by geometric features such as edges and corners, resulting in a smooth "background luminance model" image I_background that approximately reflects the temperature field distribution.

[0029] Background luminance field normalization and compensation: For each pixel (x, y) in the luminance channel, its original luminance is V_original(x,y), and the corresponding background luminance is V_bg(x,y) (from I_background). Calculate the luminance compensation coefficient Gain(x,y)=V_target / V_bg(x,y). For darker background areas (small V_bg), a larger gain (Gain>1) is needed to brighten them; for overly bright background areas (large V_bg), a smaller gain (Gain<1) is needed to darken them. Correct the luminance channel V_corrected(x,y)=V_original(x,y) * Gain(x,y).

[0030] Image reconstruction and output: The corrected luminance channel (V_corrected or L_corrected) is merged back with the original hue / saturation channel (H and S channels in HSV) or color channel (a and b channels in Lab). The image is converted from HSV / Lab color space back to RGB color space (if a color image is needed for subsequent processing), or the corrected luminance channel image is directly used as the enhanced grayscale image to obtain a uniformly bright image of the billet end face, which is provided to the downstream semantic segmentation deep learning model for end face region recognition.

[0031] The semantic segmentation unit is processed through the following steps: Model preparation and loading: Select a lightweight, high-precision semantic segmentation model suitable for industrial segmentation (such as U-Net, DeepLabv3+, etc.), collect a large number of labeled billet end face images, use the labeled data to perform end-to-end supervised training on the selected network, and after training, export the model to an optimized format (such as ONNX, TensorRT engine or OpenVINOIR) and integrate it into the vision workstation software.

[0032] Forward inference (model calculation end face region): The obtained brightness-consistent image is normalized, and the preprocessed image is input into the loaded semantic segmentation model. The model outputs a probability map or score map with the same spatial size as the input image.

[0033] Generate a binary mask: Apply a threshold (e.g., 0.5) to the probability map, if probability ≥ 0.5 then pixel = 1 (foreground); else pixel = 0 (background); resulting in a binary image (mask) where white areas (pixel value = 1) represent the "cast end face" identified by the model, and black areas (pixel value = 0) represent the background.

[0034] Post-processing and contour extraction: A light closing operation (dilation followed by erosion) is performed on the binary mask to fill in small holes that may be caused by reflection or water stains. A light opening operation (erosion followed by dilation) is then performed to smooth the edges and remove tiny isolated noise points. A contour search algorithm is used to search for connected components in the binary mask and returns a list of contour point sets. If multiple contours are found, they are filtered according to the contour area or the size of the bounding rectangle, and the largest contour is retained as the end face contour of the casting billet.

[0035] Output and Transmission: Finally, the algorithm obtains an ordered set of points. These points are connected in sequence to form a precise pixel-level contour of the billet end face. The area enclosed by the contour is the calculated end face area. The pixel area can be initially estimated by calculating the number of pixels in the contour. This contour point set is then passed to the subsequent "contour smoothing and feature extraction module" for corner detection, diagonal calculation, and true area calculation through plane calibration and perspective transformation.

[0036] Example 3: In this application, after the billet is cut and enters the conveyor roller, a billet code and flow number will be generated. After the billet is weighed, the same code and corresponding flow number will be generated. The billet steel grade, casting speed, element composition, exit temperature and billet weight will be matched to establish a complete "process fingerprint" file for each cut billet and accurately bind it to the final result (actual weight).

[0037] (1) It is conducive to realizing closed-loop feedback control (solving the problem of "weighing feedback lag and utilization") The system can match the current weighing result W through code matching. 称重 By accurately associating it with the billet from which the specific cutting command L_cut was previously issued, the system can immediately calculate the weight error: ΔW = W 定重 -W 称重 This error ΔW will be used for feedback fine-tuning: L2 = L1 + (W 定重 -W 称重 ) / (A1*ρ), thereby correcting the cutting length of the next or subsequent billets under the same conditions, and realizing closed-loop control.

[0038] (2) It is conducive to driving the continuous optimization of AI models (solving the problem of "improving model accuracy") For example, it can provide training data for a random forest density model: each set of matching records (steel grade, elemental composition, drawing speed, etc.) -> (actual weight, actual area, cutting length) is a perfect training sample. The system can periodically or online retrain the model with this new data, making the density prediction ρ increasingly accurate.

[0039] (3) Support production traceability and quality analysis When a quality problem occurs in a downstream process of a cast billet, its unique code can be used to trace back to all process parameters and visual images of the continuous casting process, facilitating root cause analysis. After accumulating a large amount of data, it is possible to analyze the distribution of the pass rate for fixed length and weight under different steel grades and process parameters, providing data support for optimizing continuous casting processes (such as casting speed and cooling intensity).

[0040] (4) Achieve synchronization of "flow" and handle multi-flow production. Multi-strand continuous casting machines produce multiple slabs simultaneously (e.g., first and second strands). The process conditions (casting speed, cooling) for each strand may differ slightly. "Strand sequence" information ensures that data and feedback are matched and transmitted correctly on the production line, preventing data from the first strand from being incorrectly used to adjust the cutting of the second strand, thus guaranteeing the independence and accuracy of multi-line production.

[0041] Example 4: This application predicts the thermal shrinkage rate R of the billet based on the billet steel grade, casting speed, elemental composition, and outlet temperature using a thermal shrinkage rate model, and then uses the formula L... 定尺切割 =L 定尺 *R calculates the fixed-length cutting length, where L 定尺 This indicates the length value that the cast billet must reach after cooling, as specified by the production plan or order.

[0042] The thermal shrinkage rate model is constructed through the following steps: Step 1: Data Preparation and Sample Construction Feature (input X) acquisition: Steel grade: A category variable that determines the matrix shrinkage characteristics of the material.

[0043] Casting speed: A numerical variable. Casting speed affects the solidification and cooling rate and the temperature field of the billet, indirectly affecting shrinkage.

[0044] Key elemental composition: numerical variables, especially C, Si, Mn, P, and S.

[0045] The carbon content is crucial: it determines the phase transformation point of steel (such as the A3 point), and the volume change accompanying the phase transformation is the main part of the shrinkage.

[0046] Si and Mn affect the stability of austenite.

[0047] P and S affect the solidification zone and segregation, thus affecting the uniformity of shrinkage.

[0048] Outlet temperature: a numerical variable, also known as the "temperature at which molten steel exits the tundish". This is the starting temperature of the billet solidification and is the basis for calculating the temperature drop process.

[0049] Label (output y) calculation: It is necessary to know the "actual thermal shrinkage rate" of each historical casting billet as a training objective.

[0050] Calculation formula: R_actual = L_cut_actual / L_final_actual Data source: L_cut_actual: The actual cutting length of the billet (obtained from the control system record).

[0051] L_final_actual: The actual length of the billet after it has cooled to the target state.

[0052] For cold feed onto the cooling bed: the target state is room temperature. L_final_actual comes from the cold length measurement data after the cooling bed.

[0053] For hot-feed rolling: the target state is the furnace inlet temperature. L_final_actual is difficult to obtain directly and may need to be estimated using a high-temperature length measuring instrument installed on the heat-insulating roller conveyor, or by using theoretical values ​​in the initial stage and combining them with the later rolling dimensions to deduce the final value.

[0054] By using a large amount of historical data, the R_actual corresponding to each casting billet is calculated, forming a label for supervised learning.

[0055] Step 2: Model selection, training, and validation Model selection: Random forest regression algorithm was selected.

[0056] It can handle mixed-type features (categorical steel grades, numerical elements, and temperature).

[0057] It can capture complex nonlinear relationships between characteristics and shrinkage rate (such as the interaction between C content and temperature on shrinkage).

[0058] The importance ranking of output features can be used to verify metallurgical principles (for example, it may be found that C content and outlet temperature are the two most important features).

[0059] Feature engineering and training: Steel grade coding: Use unique heat coding.

[0060] Feature standardization: Standardize numerical features.

[0061] Data set partitioning: Divide the dataset into training and test sets according to a set ratio.

[0062] Model training: Use the training set data (X_train, y_train) to train a random forest model and learn the mapping relationship from process parameters to shrinkage rate.

[0063] Model Validation and Storage: The model performance (RMSE, R²) is evaluated using a test set. The goal is to minimize the error in predicting the shrinkage rate R_predict, as a small error in the shrinkage rate will be directly amplified into a length error.

[0064] Save the trained model and integrate it into the online system.

[0065] Example 5: In this application, the density ρ of the billet is first calculated based on the steel grade, casting speed, and elemental composition of the billet, using a machine learning random forest algorithm (i.e., density prediction model).

[0066] Density prediction models can be constructed through the following steps: Step 1: Data Collection and Sample Library Construction Historical production data is collected, with each data record representing a produced billet, including: Feature (input X): Steel grade: Category variables (such as Q235B, 45#, 304, etc.) need to be coded.

[0067] Pulling speed: A numerical variable, usually measured in meters per minute.

[0068] Elemental composition: numerical variables, including the percentage content of major elements such as C, Si, Mn, P, and S. Note that S and P are critical characteristics and must be included.

[0069] Label (output y): Actual density of the billet (ρ_actual): This is the target value for model learning.

[0070] How to obtain ρ_actual: Calculate it by back-calculating from historical data.

[0071] ρ_actual = W_actual / (A_actual × L_cut) in: W_actual: The actual weighing weight of the billet (from the weighing system).

[0072] A_actual: The actual end face area of ​​the billet before cutting (from historical visual inspection data).

[0073] L_cut: The actual cutting length of the billet (recorded from the control system).

[0074] Step 2: Data Preprocessing Steel grade coding: Converting the text category "steel grade" into a machine-readable format. One-hot encoding is commonly used, creating a binary feature for each steel grade.

[0075] Feature standardization: Standardize numerical features (pulling speed, content of each element) (such as Z-score standardization) to make their mean 0 and variance 1, so as to improve the training efficiency and stability of the model.

[0076] Data cleaning: Remove obviously erroneous or invalid records (such as abnormal data with area or weight values ​​of 0 or negative).

[0077] Step 3: Training the Random Forest Model Model selection: Random forest regressor was chosen.

[0078] Training process: The preprocessed dataset is divided into a training set (e.g., 80%) and a test set (e.g., 20%).

[0079] On the training set, multiple sample subsets are randomly selected using a bootstrap sampling method.

[0080] Build a decision tree for each subset of samples. At the split nodes of the tree, randomly select a subset of features (such as the square root of the total number of features) to find the optimal split point to predict the density.

[0081] A "forest" is formed by training multiple trees in parallel.

[0082] Model output: The completed random forest model's prediction is the average of all decision tree predictions.

[0083] Step 4: Model Validation and Saving Performance Evaluation: The model accuracy is evaluated using a test set. Key metrics include: Root mean square error (RMSE): reflects the average error in density prediction.

[0084] Coefficient of determination (R²): reflects the model's ability to explain density fluctuations; the closer to 1, the better.

[0085] Model solidification: The trained model (including all parameters such as tree structure, split points, and leaf node values) is serialized and saved as a file (such as .pkl, .joblib, or .onnx format) and integrated into the software module of the online system.

[0086] Assume the production target is W per billet. 定重 =1000kg.

[0087] (1) The first piece (weighing billet) The measured end face area is A1 = 0.1 m². According to the formula L1 = W 定重 Calculate the required cutting length using (A1*ρ), then cut the material. After cutting, weigh the material and find the actual weight W. 称重 = 990 kg, which is 10 kg lighter than the target.

[0088] (2) System feedback and calculation System record: A1 = 0.1 m², W 称重 = 990 kg; First fine-tuning of the cutting length: L2 = L1 + (W 定重 -W 称重) / (A1*ρ)。

[0089] (3) The second piece (the blank to be cut) The vision system measures the current end face area of ​​the billet to be cut as A2 = 0.102 m² (slightly larger than the first piece, possibly due to a slight bulge); the system calculates the area scaling factor: ≈ 1.01; Final length calculation: L3 = 1.01 * (L2 + other compensation items). This calculation not only compensates for the weight deviation of the previous 10kg piece, but also takes into account the impact of a 1% increase in the cross-sectional area of ​​the current billet. By increasing the length by about 1%, it offsets the potential overweight caused by the increased area, thus more accurately hitting the 1000kg target.

[0090] Other compensation items in this application include compensation for changes in pulling speed and compensation for changes in elements. Since sulfur (S) and phosphorus (P) are "critical harmful impurity elements", in order to simplify the calculation process, the compensation for changes in elements only considers the changes in sulfur and phosphorus. Therefore, L3 = 1.01 * (L2 + ΔB1 * K1 + ΔB2 * K2 + ΔB1 * K2), where ΔB1, ΔB2, and ΔB3 represent the changes in pulling speed, sulfur, and phosphorus, respectively, and K1, K2, and K3 represent the coefficients (i.e., weights) of the changes in pulling speed, sulfur, and phosphorus, respectively.

[0091] Example 6: The casting speed variation coefficient K1, sulfur element variation coefficient K2, and phosphorus element variation coefficient K3 of this application can be obtained through historical production data mining, machine learning model training, and process calibration verification. These coefficients are strongly correlated with the steel grade of the cast billet and need to be calibrated separately for each steel grade. The specific acquisition method is as follows: (1) Construct a multi-dimensional process-weight sample library Based on the "process fingerprint" archive of the casting billet established by the matching weighing module, full-dimensional data of each casting billet in historical production are collected to form training samples. Each sample must include: Independent variables: actual casting speed, percentage of sulfur (S) content, percentage of phosphorus (P) content, steel grade of the billet, actual end face area, cutting length, and exit temperature; Dependent variables: actual weighing weight of the billet, and the deviation ΔW between the target weight and the actual weight; Related parameters: the change in drawing speed ΔB1 between the blank to be cut and the weighed blank, the change in sulfur element ΔB2, the change in phosphorus element ΔB3, and the corresponding cutting length correction ΔL (i.e. the cutting length value that needs to be adjusted due to parameter changes).

[0092] The samples need to cover different steel grades and different process parameter fluctuation ranges, and abnormal data (such as invalid data caused by weighing errors or severe end face deformation) should be removed.

[0093] (2) Inferring the initial value of the coefficient based on the error The physical meaning of coefficients K1, K2, and K3 is the cutting length correction coefficient corresponding to a unit change in parameters, that is, the cutting length value that needs to be adjusted for every unit change in pulling speed / S / P. The initial value is derived by reverse calculation using the correction formula for constant weight cutting length: The cutting length correction logic ΔL = ΔB1*K1 + ΔB2*K2 + ΔB1*K2 can be transformed into K = (ΔL) / (ΔB), where: For K1: K1 = ΔL 拉速 / ΔB1 (When only the pulling speed changes and other parameters remain unchanged, K1 = the cutting length correction caused by the pulling speed change ÷ the pulling speed change). For K2: K2 = ΔL S / ΔB2 (When only the S element changes and other parameters remain unchanged, K2 = the cutting length correction caused by the change in S ÷ the change in the S element). For K3: K3 = ΔL P / ΔB3 (When only P element changes and other parameters remain unchanged, K3 = cutting length correction caused by P change ÷ P element change).

[0094] By screening historical samples where a single parameter fluctuates while other parameters remain stable, the cutting length correction corresponding to the unit change of each parameter is calculated to obtain the initial empirical values ​​of K1, K2, and K3.

[0095] (3) Accurate coefficient calibration based on random forest regression The initial empirical values ​​are optimized using machine learning to achieve accurate coefficient calibration. The steps are as follows: Step 1: Feature and Label Definition; Input characteristics: billet steel grade, ΔB1, ΔB2, ΔB3, end face area ratio (A2 / A3); Output label: Actual cutting length correction ΔL (i.e., the actual value of the cutting length that needs to be adjusted due to parameter changes in order to match the target weight in actual production).

[0096] Step 2: Model training; The random forest regression algorithm is used to train the sample database and learn the nonlinear mapping relationship between parameter changes and cutting length correction. After the model is trained, the contribution weights of each parameter (ΔB1, ΔB2, ΔB3) to ΔL are output through feature importance analysis. These weights are the optimized K1, K2, and K3.

[0097] Step 3: Model Validation; The model accuracy was verified using a test set, with a weight deviation of ±1.5% as the passing standard. If the verification results did not meet the standard, additional samples were added and the model was retrained until the cutting length correction corresponding to the coefficients could achieve a weight passing rate of over 90%.

[0098] (4) Establish a coefficient database according to steel type Because the metallurgical properties (such as solidification shrinkage, density, and element sensitivity) of different steel grades vary significantly, K1, K2, and K3 need to be calibrated separately according to the steel grade. For example, low carbon steel is more sensitive to changes in drawing speed, so its K1 value is larger. Low-alloy steel is more sensitive to changes in the content of S and P elements, and has larger K2 and K3 values. By dividing the sample library according to steel type, training the model separately, and establishing a steel type-coefficient (K1, K2, K3) correspondence library, the system can automatically match the corresponding coefficient according to the current billet steel type when running.

[0099] (5) Continuous optimization of coefficients based on closed-loop feedback During system operation, each parameter change (ΔB1, ΔB2, ΔB3), cutting length correction (ΔL), and actual weight deviation are used as new samples and continuously added to the sample library. The random forest model is retrained periodically with new samples to update K1, K2, and K3. When a decrease in the weight qualification rate or a significant adjustment in process parameters (such as changes in the drawing speed range or optimization of steel grade processes) is detected, the model is retrained immediately to ensure that the coefficients are adapted to the latest production process.

[0100] It should also be noted that this application considers the impact of changes in S and P elements, rather than C, Si, and Mn, which have higher content. The main reason is that C, Si, and Mn, as the main additive elements, have their content fluctuations strictly controlled during production after the steel grade is determined. Significant changes are rare, and they are considered steady-state process parameters, rather than process fluctuation parameters that require real-time dynamic correction. Although the content of S and P is relatively low, in the refining stage of continuous casting production, the removal of S and P is affected by the stability of desulfurization and dephosphorization processes, and is prone to small fluctuations. Moreover, the presence of S and P can directly lead to defects such as hot brittleness and cold brittleness in the billet, and they are key impurity indicators that are monitored online in on-site production. This application includes S and P in the dynamic correction based on the actual monitoring needs of on-site process fluctuations, rather than simply considering the weight of the element's influence on density.

[0101] In summary, this application has the following significant advantages over the prior art: (1) Significantly improves the accuracy of weighing and reduces material waste. By visually measuring the geometry of each casting billet end face and combining it with a dynamic density compensation algorithm, the volume calculation error caused by defects such as squareness and bulging in traditional methods is effectively overcome, so that the weight qualification rate reaches more than 90% (weight less than ±1.5%). Based on closed-loop feedback control of multi-sensor data fusion (weighing, temperature, casting speed, elemental composition) and combined with the designed model, the cutting length is adaptively adjusted, reducing head and tail losses and improving the yield.

[0102] (2) Fully automated operation, reducing human intervention The system automatically calculates the optimal cutting length, completely avoiding inconsistencies caused by manual adjustments based on experience. Through an online learning mechanism, it continuously optimizes model parameters to adapt to changes in different steel grades and processes, achieving "unmanned" precise control.

[0103] (3) Intelligent anomaly handling capability When severe deformation (such as delamination or excessive bulging) or abnormal weight during production is detected, the system can trigger an alarm and recommend corrective strategies to prevent the generation of continuous defects and improve product quality consistency.

[0104] (4) Data-driven continuous optimization Establish a full lifecycle process database, conduct historical data backtesting and analysis, explore the correlation between cutting parameters and weight error, and continuously optimize cutting parameters.

[0105] (5) Reduce costs Precise weight control can reduce the pass rate of fixed length and weight, and improve the yield rate; no manual operation is required during production, thus reducing labor costs.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A vision-based online length and weight determination system for continuously cast billets, characterized in that, include: The image acquisition module is used to acquire images of the end face of the cast billet; The end face detection module is used to process the end face image, extract the contour and detect defects such as squareness and bulging, and calculate the actual end face area. The data synchronization module is used to synchronize the values ​​of various factors of the billet, including the billet steel grade, casting speed, elemental composition, and outlet temperature. The matching weighing module is used to match the weighing information with the corresponding billet data information; The fixed-length cutting calculation module is used to calculate the fixed-length cutting based on the heat shrinkage rate; The thermal shrinkage rate R of the billet and the fixed-length cutting length L are predicted based on the billet steel grade, casting speed, elemental composition, and exit temperature. 定尺切割 =L 定尺 *R, where L 定尺 This indicates the length value that the cooled billet must reach, as specified by the production plan or order. The constant weight cutting length calculation module is used to calculate and dynamically correct the cutting length based on the target weight, end face area, and density. The fixed-weight cutting length calculation module performs the calculation through the following steps: When the billet weight is not received, based on the billet steel grade, casting speed, and elemental composition, the density ρ of the billet is calculated using the random forest algorithm, and then the fixed weight W of the billet is calculated on-site. 定重 The initial cutting length L1 of the billet is calculated based on the actual end face area A1 measured in real time using vision, and L1 = W. 定重 / (A1*ρ); After receiving the weight of the previously cut billet, the cutting length of the uncut billet is fine-tuned. The fine-tuned cutting length is L2, L2=L1+(W 定重 -W 称重 ) / (A1*ρ; If the drawing speed, end face area, and element composition of the blank to be cut and the blank to be weighed change, the cutting length should be readjusted based on the drawing speed change coefficient and the element change coefficient. The readjusted cutting length is L3. Where A2 represents the end face area of ​​the billet to be cut, A3 represents the end face area of ​​the weighed billet used as a reference, ΔB1, ΔB2, ΔB3 represent the changes in casting speed, sulfur content, and phosphorus content, respectively, and K1, K2, K3 represent the coefficients of change in casting speed, sulfur content, and phosphorus content, respectively. The display module is used to display the detection and calculation results in real time.

2. The vision-based online length and weight determination system for continuously cast billets according to claim 1, characterized in that, The image acquisition module includes an end face camera. When the laser temperature sensor detects that the billet to be cut has reached the designated position, the PLC sends the arrival signal to the vision server, and the vision server activates the end face camera to take a picture.

3. The vision-based online length and weight determination system for continuously cast billets according to claim 1, characterized in that, The end face detection module includes an image preprocessing unit, a semantic segmentation unit, a contour extraction unit, and an area calculation unit. When the billet travels to the designated position in the image, the end face detection module performs end face detection. The image preprocessing unit preprocesses the brightness of the billet end face image, converting it into an image with uniform brightness. The semantic segmentation unit calculates the image region where the billet end face image is located, and the contour extraction unit extracts the edge contour of the image region. The area calculation unit performs perspective transformation on the image region through planar calibration, so that each pixel represents the same area after the transformation. The number of pixels in the transformed image region is calculated to calculate the actual area value of the billet end face.

4. The vision-based online length and weight determination system for continuously cast billets according to claim 3, characterized in that, After extracting the edge contour of the image area, contour smoothing is used to remove contour burrs. The four corner points and the center line of the contour are found. The difference between the two diagonals of the contour is used to determine whether the end face of the billet is distorted. The difference between the center line of the contour and the center line of the set center line threshold is used to determine whether the end face of the billet is bulging.

5. The vision-based online length and weight determination system for continuously cast billets according to claim 1, characterized in that, The billet steel grade refers to the type of billet, the casting speed refers to the speed at which the billet exits the crystallizer, the elemental composition refers to the elements of the molten steel after refining, and the exit temperature refers to the furnace temperature when the molten steel exits the tundish.

6. The vision-based online length and weight determination system for continuously cast billets according to claim 1, characterized in that, After the billet is cut and enters the conveyor roller, a billet code and flow number will be generated. After the billet is weighed, the same code and corresponding flow number will be generated. The billet steel grade, casting speed, element composition, exit temperature and billet weight will be matched.

7. The vision-based online length and weight determination system for continuously cast billets according to claim 1, characterized in that, The system also includes an alarm module, which is used to trigger an alarm when severe end face deformation or abnormal weight is detected.

Images