A slab longitudinal crack prediction method based on uneven state of crystallizer flow field

Abstract

The present application relates to the slab continuous casting production technical field, it is a kind of slab longitudinal crack prediction method based on mould flow field unbalanced state, comprising: collecting mould flow field characteristic parameter, billet shell solidification characteristic parameter and continuous casting process parameter, and construct operating parameter dataset;Operating parameter dataset is carried out data preprocessing, and is carried out normalization processing;From operating parameter dataset, the flow field characteristic that leads to longitudinal crack is filtered;The flow field characteristic that leads to longitudinal crack is weighted using analytic hierarchy process, entropy weight method and process correlation degree method, and combination weight is obtained;According to combination weight, longitudinal crack correlation flow field unbalanced index is calculated, and longitudinal crack correlation flow field unbalanced degree is divided;Longitudinal crack prediction model is constructed, and the output result of longitudinal crack correlation flow field unbalanced index and longitudinal crack prediction model is combined to grade forecast longitudinal crack.

Description

Technical Field

[0001] This invention relates to the field of slab continuous casting production technology, and in particular to a method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer. Background Technology

[0002] Longitudinal cracks in slabs are a common and serious surface defect in continuous casting production. They are mainly distributed in the center of the wide face or near the narrow face of the slab, with crack depths ranging from several millimeters to tens of millimeters. This not only significantly increases the cost of slab re-grinding but can also lead to slab scrapping in severe cases. Furthermore, crack propagation during subsequent rolling can cause steel plate fracture, directly affecting product mechanical properties and production safety. The core formation mechanism of longitudinal cracks in slabs is directly related to the flow field of molten steel within the crystallizer: an uneven flow field leads to uneven solidification rates of the slab shell, insufficient meniscus stability, and localized scouring and thinning of the slab shell. Simultaneously, it causes stress concentration at the solidification front. When the stress exceeds the high-temperature strength of the slab shell, longitudinal cracks will occur. Existing slab longitudinal crack prediction technologies mostly focus on process parameters or single solidification characteristics, failing to establish a precise correlation between flow field imbalance and longitudinal cracks, exhibiting significant technical shortcomings, as detailed below: The experience-based judgment method relies on operators to observe the liquid level in the crystallizer on-site and adjust process parameters based on experience. It is highly subjective, requires high skill levels from operators, cannot predict the risk of longitudinal cracks in advance, and is mostly handled after the fact.

[0003] Single-parameter monitoring method: Longitudinal cracks are predicted by using single parameters such as thermocouples to monitor the temperature of the crystallizer wall and ultrasonic instruments to monitor liquid surface fluctuations. This method does not consider the comprehensive influence of the flow field on the solidification of the billet shell. For example, monitoring only the liquid surface fluctuations without linking them to the temperature gradient at the solidification front leads to one-sided predictions and a high rate of misjudgment and missed judgment.

[0004] Numerical simulation method: CFD model is used to simulate the flow field and solidification process, which can analyze the influence of the flow field on the billet shell. However, the calculation is large and a single set of working conditions simulation takes 8-36 hours, which cannot meet the real-time requirements of industrial online forecasting. In addition, the simulation results deviate from the actual working conditions, lack actual data verification, and the forecast accuracy is limited.

[0005] Traditional machine learning prediction methods: Some studies use a single neural network model to predict longitudinal cracks with process parameters as input. They do not incorporate the core features of flow field imbalance and lack constraints on billet solidification dynamics. They have weak generalization ability when faced with changes in parameters such as casting speed and superheat, poor adaptability across working conditions, low actual prediction accuracy, and cannot achieve graded prediction.

[0006] Blind spots in core technologies: Existing technologies have not constructed a correlation system between the flow field imbalance state, the shell growth characteristics, and the formation of longitudinal cracks. They have not clearly defined the flow field characteristic indicators specific to longitudinal cracks, and cannot accurately quantify the inducing effect of flow field imbalance on longitudinal cracks, resulting in a lack of core theoretical support for prediction.

[0007] In summary, existing technologies suffer from several key limitations: the lack of a dedicated quantitative system for quantifying flow field imbalances specific to longitudinal cracks, the absence of solidification physics constraints in the prediction models, the lack of graded prediction and precise control mechanisms, and insufficient online real-time prediction capabilities. Therefore, developing a graded prediction method for slab longitudinal cracks based on the flow field imbalance state in the crystallizer, adaptable to online industrial production, has become an urgent need for improving the quality and efficiency of continuous casting production. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer. This invention is based on the mechanism that longitudinal cracks in slabs are caused by uneven solidification and stress concentration in the slab shell due to unbalanced flow field. It constructs a complete process architecture encompassing multi-source data acquisition, data preprocessing, quantification of flow field imbalance specific to longitudinal cracks, solidification-constrained CNN-GRU prediction, graded early warning, and closed-loop control. The core innovations lie in three aspects: the quantification system of flow field imbalance specific to longitudinal cracks, the fusion of solidification dynamics constraints and deep learning models, and graded prediction.

[0009] The technical means employed in this invention are as follows: A method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer includes: collecting flow field characteristic parameters of the crystallizer, solidification characteristic parameters of the slab shell, and continuous casting process parameters, and constructing an operating parameter dataset; preprocessing and normalizing the operating parameter dataset; filtering flow field characteristics that lead to longitudinal cracks from the operating parameter dataset; assigning weights to the flow field characteristics leading to longitudinal cracks using the analytic hierarchy process (AHP), entropy weighting, and process correlation method, and obtaining combined weights; calculating the longitudinal crack-related flow field imbalance index based on the combined weights, and classifying the degree of longitudinal crack-related flow field imbalance; constructing a longitudinal crack prediction model, and training and optimizing the longitudinal crack prediction model; and performing graded prediction of longitudinal cracks by combining the longitudinal crack-related flow field imbalance index and the output results of the longitudinal crack prediction model.

[0010] Furthermore, the flow field characteristic parameters of the crystallizer include: meniscus molten steel flow velocity, crystallizer liquid surface fluctuation characteristics, narrow-face molten steel backflow intensity, and molten steel impact characteristics at the nozzle; the solidification characteristic parameters of the billet shell include: solidification front temperature gradient, billet shell thickness distribution characteristics, and meniscus billet shell solidification-melting cycle frequency; the continuous casting process parameters include: billet casting speed, submerged nozzle immersion depth, molten steel superheat, crystallizer vibration parameters, and protective slag viscosity.

[0011] Furthermore, the data preprocessing includes: noise removal, missing value imputation, spatiotemporal dimension alignment, and outlier removal; The noise removal method refers to using a combination of wavelet transform and moving median averaging for random noise removal to eliminate high-frequency interference and burst noise; and using Gaussian filtering for high-speed camera image noise removal to preserve meniscus flow edge features. The missing value imputation refers to the imputation of missing values ​​caused by temporary sensor failures. Short-term missing values ​​are imputed by linear interpolation, while long-term missing values ​​are imputed by the weighted average of historical data under the same operating conditions to ensure data continuity. The spatiotemporal alignment means that, based on the 10Hz solidification front temperature parameter timestamp, the flow field characteristic parameters are downsampled to 10Hz, and the process parameters are linearly interpolated to 10Hz, with a unified sampling frequency of 10Hz, to ensure data integrity within the same time step. The outlier removal means using... The criteria combine process boundary constraints to eliminate outlier data. When data exceeds the mean standard deviation or the reasonable range of the process, it is judged as an outlier and replaced with the mean of the same operating conditions. The normalization process maps all feature parameters to the [0,1] interval, eliminates dimensional differences, and uses the min-max normalization formula:

[0012] in, The original data, This represents the normalized data; , These represent the minimum and maximum values ​​of this parameter under the same operating conditions.

[0013] Furthermore, the flow field characteristics leading to longitudinal cracks include: uniformity of molten steel flow velocity at the meniscus, stability of liquid surface fluctuations, uniformity of temperature gradient at the solidification front, difference in backflow intensity of molten steel at the narrow face, and offset of nozzle impact position, wherein: The uniformity of the molten steel flow velocity on the meniscus The formula used to represent the difference in flow velocity on both sides of the meniscus is:

[0014] in, , The flow velocity on both sides of the meniscus is the flow velocity. The value range is 0-1; Liquid level fluctuation stability The coefficient of variation, used to represent the amplitude of liquid level fluctuations in the crystallizer, is calculated using the following formula:

[0015] in, Let A be the standard deviation of the liquid level fluctuation amplitude, and let A be the mean of the liquid level fluctuation amplitude. The value is greater than 0; The uniformity of the temperature gradient at the solidification front The formula used to represent the difference in temperature gradient at the solidification front at the same depth on both sides of the wide face is:

[0016] in, , The average temperature gradient on both sides of the wide surface is represented by F3, which ranges from 0 to 1. The narrow-faced molten steel reflow intensity difference The formula used to represent the difference in the intensity of molten steel reflow on both narrow sides is as follows:

[0017] in, , For the backflow intensity on both narrow sides, the value of F4 is greater than 0; The offset of the water inlet impact position The formula used to indicate the deviation between the actual impact position and the ideal impact position of molten steel at the nozzle is as follows:

[0018] Where L represents the actual impact location. For the ideal impact location, the value of F5 is greater than 0.

[0019] Furthermore, the analytic hierarchy process (AHP) constructs a judgment matrix based on the continuous casting process and the longitudinal crack mechanism to determine the subjective weights of the flow field characteristics that lead to longitudinal cracks. The entropy weight method, based on historical production data, calculates the information entropy of each feature to determine the objective weights of the flow field features that lead to longitudinal cracks. The process correlation method determines the correlation weight of the flow field characteristics that lead to the longitudinal crack by calculating the correlation between each feature and the longitudinal crack. ; The combined weights Represented as:

[0020] in, These are the weighting coefficients corresponding to the subjective weights. This refers to the weighting coefficient corresponding to the objective weight.

[0021] Furthermore, the longitudinal crack-related flow field imbalance index LFI is calculated based on the combined weights and the normalized eigenvalues, and the calculation formula is as follows:

[0022] in, For the normalized first 1 eigenvalue, For the first The combined weights corresponding to the flow field characteristics that lead to longitudinal cracks; Based on the value of the longitudinal crack-associated flow field imbalance index (LFI) and historical longitudinal crack occurrence data, the degree of flow field imbalance is divided into three categories: balanced, unbalanced, and severely unbalanced. When 0 ≤ LFI < 0.25, the risk of longitudinal cracks is low, which is the balanced state. When 0.25 ≤ LFI < 0.6, the risk of longitudinal cracks is high, which is the unbalanced state. When LFI ≥ 0.6, the risk of longitudinal cracks is extremely high, which is the severely unbalanced state.

[0023] Furthermore, the longitudinal crack prediction model outputs longitudinal crack prediction results by embedding solidification dynamics constraints and process solidification coupling constraints of the billet shell, and combining the spatial feature extraction capability of CNN with the temporal feature capture capability of GRU. The embedded billet shell solidification dynamics constraint is based on the solidification dynamics principle of continuous casting billet shell, and adds the following constraints: meniscus molten steel flow velocity M≤0.8m / s; solidification front temperature gradient N≤8℃ / mm; billet shell critical thickness O≥3mm. When the input data violates the embedded billet shell solidification dynamics constraint, it is directly marked as high risk and a warning is triggered first. The process solidification coupling constraint embeds the correlation between process parameters and solidification characteristics into the longitudinal crack prediction model. Increasing the casting speed will lead to a shorter solidification time for the billet shell, and higher superheat will reduce the strength of the billet shell. This guides the model to learn the coupling relationship and improves its adaptability across working conditions.

[0024] Furthermore, the graded forecast combines the longitudinal crack-related flow field imbalance index (LFI) with the output results of the longitudinal crack forecasting model, and refers to the longitudinal crack formation and propagation laws to classify the longitudinal crack risk into mild longitudinal crack risk, moderate longitudinal crack risk and severe longitudinal crack risk. The risk of mild longitudinal cracks is defined as follows: when 0.25≤LFI<0.4, the model predicts a longitudinal crack probability of 60%-80%, the flow velocity on the meniscus is slightly uneven, the liquid level fluctuates slightly, the thickness of the billet shell is slightly different, there are no obvious cracks, and there is only a potential risk of cracking; the corresponding inducing factors are slight fluctuations in the casting speed and unstable argon flow rate. The moderate longitudinal crack risk refers to a situation where 0.4 ≤ LFI < 0.6, the model predicts a longitudinal crack probability of 80%-95%, the meniscus flow velocity is significantly uneven, the liquid surface fluctuates violently, the temperature gradient at the solidification front is large, the shell thickness difference is ≥ 0.5 mm, and microcracks have appeared but have not propagated. The corresponding causes are improper nozzle immersion depth, excessively high casting speed, and excessively high superheat. The severe longitudinal crack risk indicates that when LFI≥0.6 and the model predicts a longitudinal crack probability≥95%, the meniscus flow velocity is severely uneven, the liquid level fluctuates significantly, the nozzle impact position shifts, the billet shell becomes locally thinner, microcracks have expanded, and obvious longitudinal cracks are easily formed, which may even lead to steel leakage. The corresponding inducing causes are nozzle blockage, sudden changes in casting speed, and abnormal crystallizer vibration parameters.

[0025] Compared with the prior art, the present invention has the following advantages: This invention constructs a 5-dimensional flow field imbalance feature system for longitudinal cracks, clarifies the correlation between flow field imbalance and longitudinal cracks, and achieves a quantitative accuracy of ≥96% for the longitudinal crack-related flow field imbalance index (LFI). This solves the problem of traditional technologies lacking core quantitative indicators and provides a reliable basis for longitudinal crack prediction. This invention integrates the solidification dynamics constraints of the billet shell with the CNN-GRU model, capturing both the spatial distribution characteristics of the flow field and the temporal evolution trend. The overall accuracy of longitudinal crack prediction is greater than 98.2%, and the accuracy of prediction for mild, moderate and severe risks is greater than 97%, 98.5% and 99.3%, respectively.

[0026] The severe longitudinal crack risk warning in this invention can prevent steel leakage accidents in advance, avoid economic losses for each heat of molten steel, reduce downtime for maintenance, and improve production continuity. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the slab longitudinal crack prediction method based on the unbalanced flow field state in the crystallizer in this invention.

[0029] Figure 2 This is a schematic diagram of the longitudinal crack prediction model structure in this invention.

[0030] Figure 3 This is a structural diagram of the slab longitudinal crack prediction system based on the unbalanced flow field state in the crystallizer, as described in an embodiment of the present invention. Detailed Implementation

[0031] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0034] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0035] like Figure 1 As shown, this invention provides a method for predicting longitudinal cracks in slabs based on the unbalanced flow field state in the crystallizer. The method includes: collecting characteristic parameters of the crystallizer flow field, solidification characteristic parameters of the slab shell, and continuous casting process parameters, and constructing an operational parameter dataset. Specifically, in a preferred embodiment of this invention, the characteristic parameters of the crystallizer flow field include: meniscus molten steel flow velocity, crystallizer surface fluctuation characteristics, narrow-face molten steel backflow intensity, and nozzle molten steel impact characteristics. The solidification characteristic parameters of the slab shell include: solidification front temperature gradient, slab shell thickness distribution characteristics, and meniscus slab shell solidification-melting cycle frequency. The continuous casting process parameters are used to embed the solidification dynamics constraints of the slab shell and correct the correlation between flow field characteristics and longitudinal cracks, including: slab casting speed, submerged nozzle immersion depth, molten steel superheat, crystallizer vibration parameters, and mold flux viscosity.

[0036] During implementation, the meniscus molten steel flow velocity represents the radial flow velocity of molten steel at corresponding positions on both sides of the wide face, with a measurement range of 0-1.2 m / s; the crystallizer surface ripple characteristics are used to characterize the amplitude and frequency of surface ripples; the narrow face molten steel backflow intensity characterizes the backflow intensity of molten steel in the middle of the inner wall of the narrow face on both sides of the crystallizer, with a measurement range of 0-0.8 m / s; the nozzle molten steel impact characteristics are used to characterize the actual impact position and intensity of the nozzle molten steel. The solidification front temperature gradient represents the temperature gradient of molten steel at different depths on both sides of the wide face of the crystallizer, with a measurement range of 1350-1580℃; the billet shell thickness distribution characteristics represent the real-time thickness of the billet shell at different positions in the crystallizer; the meniscus billet shell solidification-melting cycle frequency characterizes the frequency of repeated solidification and melting of the billet shell. The measurement range for billet casting speed is 0.8-1.2 m / min; the measurement range for immersion depth of the submerged entry nozzle is 120-170 mm; the measurement range for molten steel superheat is 18-45℃; the measurement range for crystallizer vibration parameters is amplitude ±2.5~±4 mm, frequency 90-260 times / min; and the measurement range for protective slag viscosity is 0.1-0.3 Pa. s.

[0037] The runtime parameter dataset undergoes data preprocessing and normalization. Specifically, in a preferred embodiment of this invention, data preprocessing includes: noise removal, missing value imputation, spatiotemporal alignment, and outlier removal. Noise removal involves using a combination of wavelet transform and moving median averaging for random noise, selecting db5 as the wavelet basis, using 6 decomposition levels, and setting the sliding window size to 12 to remove high-frequency interference and burst noise. For high-speed camera image noise, Gaussian filtering is used to preserve meniscus flow edge features. Missing value imputation... For missing values ​​caused by temporary sensor malfunctions, short-term missing values ​​(≤8 sampling points) are filled using linear interpolation, while long-term missing values ​​(>8 sampling points) are filled using the weighted average of historical data under the same operating conditions, ensuring data continuity. Spatiotemporal alignment is achieved by using a 10Hz solidification front temperature parameter timestamp as a baseline, downsampling flow field characteristic parameters (15Hz) to 10Hz, and linearly interpolating process parameters (1Hz) to 10Hz, unifying the sampling frequency to 10Hz to ensure data integrity within the same time step. Outlier removal is achieved using… The criteria combine process boundary constraints to eliminate outlier data. When data exceeds the mean ± 3 times the standard deviation or the reasonable range of the process, it is judged as an outlier and replaced with the mean under the same operating conditions to avoid interfering with subsequent calculations.

[0038] Normalization means mapping all feature parameters to the [0,1] interval to eliminate dimensional differences, using the min-max normalization formula:

[0039] in, The original data, This represents the normalized data; , These are the minimum and maximum values ​​of the parameter under the same working conditions, ensuring consistency between subsequent quantization and model calculations.

[0040] The flow field characteristics leading to longitudinal cracks are screened from the operational parameter dataset. Specifically, in a preferred embodiment of this invention, the flow field characteristics leading to longitudinal cracks include: uniformity of molten steel flow velocity at the meniscus, stability of surface fluctuations, uniformity of temperature gradient at the solidification front, difference in backflow intensity of molten steel at the narrow face, and offset of the nozzle impact position, wherein: Uniformity of molten steel flow velocity on the meniscus This is used to represent the difference in flow velocity on both sides of the meniscus. Uneven flow velocity leads to different degrees of erosion on both sides of the billet shell, resulting in a large difference in thickness, which is the core cause of longitudinal cracks. The calculation formula is:

[0041] in, , The flow velocity on both sides of the meniscus is the flow velocity. The value ranges from 0 to 1; the closer it is to 1, the more uniform it is.

[0042] Liquid surface fluctuation stability The coefficient of variation is used to represent the amplitude of liquid level fluctuations in the crystallizer. Severe fluctuations can lead to repeated solidification and melting of the billet shell, resulting in decreased strength. The calculation formula is:

[0043] in, Let A be the standard deviation of the liquid level fluctuation amplitude, and let A be the mean of the liquid level fluctuation amplitude. The value is greater than 0, and the closer it is to 0, the more stable it is. Uniformity of temperature gradient at the solidification front This is used to represent the difference in temperature gradient at the solidification front at the same depth on both sides of the wide face. Uneven gradient leads to differences in the solidification rate of the billet shell, causing stress concentration. The calculation formula is:

[0044] in, , The average temperature gradient on both sides of the wide surface is represented by F3, which ranges from 0 to 1. Narrow-faced molten steel reflow strength is poor This value represents the difference in molten steel reflow intensity between the two narrow sides. Uneven reflow leads to uneven growth of the billet shell on the narrow side, inducing longitudinal cracks at the edges. The calculation formula is:

[0045] in, , For the backflow intensity on both narrow sides, the value of F4 is greater than 0, and the larger the value, the more uneven the flow. Sprue impact position offset This is used to indicate the degree of deviation between the actual impact position of molten steel at the nozzle and the ideal impact position. This deviation can lead to localized thinning of the billet shell due to erosion, inducing localized longitudinal cracks. The calculation formula is as follows:

[0046] Where L represents the actual impact location. For the ideal impact position, the value of F5 is greater than 0, and the larger the value, the more severe the deviation.

[0047] The analytic hierarchy process (AHP), entropy weighting method, and process correlation method are used to assign weights to the flow field characteristics leading to longitudinal cracks, and a combined weight is obtained. In a preferred embodiment of this invention, the AHP constructs a judgment matrix based on the continuous casting process and the mechanism of longitudinal cracks to determine the subjective weights of the flow field characteristics leading to longitudinal cracks. The entropy weight method calculates the information entropy of each feature based on historical production data, and determines the objective weights of the flow field features that lead to longitudinal cracks. The process correlation method determines the correlation weight of the flow field characteristics that lead to the longitudinal crack by calculating the correlation between each feature and the longitudinal crack. ; Combine weights Represented as:

[0048] in, These are the weighting coefficients corresponding to the subjective weights. This refers to the weighting coefficient corresponding to the objective weight. Take 0.3, Taking 0.3, the final combined weight is obtained. This ensures that the weights are tilted toward features that have a greater impact on longitudinal cracks.

[0049] The longitudinal crack-related flow field imbalance index (LFI) is calculated based on the combined weights, and the degree of imbalance in the longitudinal crack-related flow field is classified. In a preferred embodiment of the present invention, the longitudinal crack-related flow field imbalance index (LFI) is calculated based on the combined weights and the normalized eigenvalues, and the calculation formula is as follows:

[0050] in, For the normalized first 1 eigenvalue, For the first The combined weights corresponding to the flow field characteristics that lead to longitudinal cracks; F1 and F3 are uniformity indices, and the larger the value after normalization, the better; F2, F4, and F5 are imbalance indices, and the larger the value after normalization, the worse. Before calculation, they were unified to the logic that the larger the value, the more unbalanced the flow field.

[0051] Based on the value of the longitudinal crack-associated flow field imbalance index (LFI) and historical longitudinal crack occurrence data, the degree of flow field imbalance is divided into three categories: equilibrium, imbalance, and severe imbalance. When 0 ≤ LFI < 0.25, the risk of longitudinal cracks is low, indicating an equilibrium state. When 0.25 ≤ LFI < 0.6, the risk of longitudinal cracks is high, indicating an imbalance state. When LFI ≥ 0.6, the risk of longitudinal cracks is extremely high, indicating a severe imbalance state.

[0052] A longitudinal crack prediction model is constructed, and the longitudinal crack prediction model is trained and optimized. In a specific implementation, as a preferred embodiment of the present invention, the longitudinal crack prediction model outputs longitudinal crack prediction results by embedding the solidification dynamics constraints of the billet shell and the process solidification coupling constraints, and combining the spatial feature extraction capability of CNN and the temporal feature capture capability of GRU. The embedded billet shell solidification kinetics constraint (hard constraint) is based on the solidification kinetics principle of continuous casting billet shell and adds the following constraints: excessive flow velocity will erode the billet shell, resulting in insufficient billet shell thickness, so the flow velocity of molten steel at the meniscus is limited to M≤0.8m / s; excessive gradient will cause uneven billet shell shrinkage and stress concentration, so the solidification front temperature gradient N≤8℃ / mm is limited; insufficient billet shell thickness at the crystallizer outlet is prone to cracking, so the critical billet shell thickness O≥3mm is limited. When the input data violates the embedded billet shell solidification kinetics constraint, it is directly marked as high risk and a warning is triggered first. The process solidification coupling constraint (soft constraint) embeds the correlation between process parameters and solidification characteristics into the longitudinal crack prediction model. Increasing the casting speed will lead to a shorter solidification time of the billet shell, and higher superheat will reduce the strength of the billet shell. This guides the model to learn the coupling relationship, improves the adaptability across working conditions, and avoids prediction distortion when process parameters change.

[0053] During implementation, two types of specific constraints are embedded in the model input and training process to align with the longitudinal crack formation mechanism, improve the model's generalization ability and prediction rationality, and avoid misjudgments that do not conform to the solidification law. For example... Figure 2 As shown, the longitudinal crack prediction model integrates the spatial feature extraction capabilities of CNN to capture the spatial distribution differences of flow field and solidification features, such as the temperature gradient distribution on both sides of the wide face, and the temporal feature capture capabilities of GRU to capture the evolution trend of flow field imbalance, such as the process of liquid surface fluctuation from stable to violent. GRU has a simpler structure and faster computation than LSTM, making it suitable for real-time prediction. Combined with the channel attention mechanism, it improves the utilization rate of core features. The model structure consists of 5 layers: Input layer: The input dimension is (time step T, feature dimension D), T is 40 (corresponding to 4s data, covering the evolution process of uneven flow field), D is 13 (5 longitudinal crack-specific flow field features, 5 process parameters, and 3 solidification features), and the input data is normalized spatiotemporally aligned data.

[0054] The CNN spatial feature extraction layer contains 3 convolutional layers and 2 max pooling layers. The convolutional kernel size is 3×3, the stride is 1, the padding method is the same, the activation function is ReLU, the pooling kernel size is 2×2, and the stride is 2. It extracts the spatial distribution patterns of flow field and solidification features (such as the spatial differences in flow velocity and temperature gradient on both sides), providing spatial dimension support for forecasting.

[0055] GRU temporal feature extraction layer: contains 2 GRU layers, each with 156 hidden neurons and a dropout rate of 0.25. It captures the temporal variation trend of flow field imbalance, reflects the continuous impact of flow field imbalance on the billet shell, and solves the shortcomings of traditional models that can only capture single-point states. Channel attention fusion layer: The spatial features extracted by CNN and the temporal features extracted by GRU are weighted and fused to highlight the key features (such as meniscus flow velocity and temperature gradient) and key time steps (such as the time point of LFI mutation) that affect longitudinal cracks, thereby improving the prediction accuracy.

[0056] Output layer: The softmax activation function is used to output the probability distribution of three types of longitudinal crack risk levels (mild, moderate and severe). The category with the highest probability is the final prediction result. At the same time, the LFI value is output to provide a basis for process control.

[0057] The longitudinal crack prediction model was trained and optimized as follows: Dataset construction: Collect industrial field data (600 sets), label the risk level of longitudinal cracks (no risk, mild, moderate, severe) based on the surface inspection results of the billet, and divide it into training set (70%), validation set (20%), and test set (10%). Training parameters: AdamW optimizer, learning rate 0.0008, batch size 48, number of iterations 220, cross-entropy loss function, L2 regularization (weight decay coefficient 0.0012) added to prevent overfitting, and early stopping method used to avoid overtraining of the model. Model optimization: Bayesian optimization was used to adjust hyperparameters such as the number of convolutional kernels and the number of neurons in the GRU hidden layer, and the optimal structure was finally determined to ensure that the prediction accuracy of the model on the test set is ≥98.2%, and the prediction accuracy of each level is ≥97%, which meets the needs of industrial online real-time forecasting.

[0058] The longitudinal crack is graded and predicted by combining the longitudinal crack-related flow field imbalance index (LFI) and the output of the longitudinal crack prediction model. In a preferred embodiment of the invention, the graded prediction combines the longitudinal crack-related flow field imbalance index (LFI) and the output of the longitudinal crack prediction model, and, with reference to the formation and propagation laws of longitudinal cracks, classifies the longitudinal crack risk into mild longitudinal crack risk, moderate longitudinal crack risk, and severe longitudinal crack risk. The risk of mild longitudinal cracks is defined as follows: when 0.25≤LFI<0.4, the model predicts a longitudinal crack probability of 60%-80%, the flow velocity on the meniscus is slightly uneven, the liquid level fluctuates slightly, the thickness of the billet shell is slightly different, there are no obvious cracks, and there is only a potential risk of cracking; the corresponding inducing factors are slight fluctuations in the casting speed and unstable argon flow rate. A moderate risk of longitudinal cracks indicates that when 0.4 ≤ LFI < 0.6, the model predicts a longitudinal crack probability of 80%-95%, resulting in significant uneven flow velocity at the meniscus, violent fluctuations in the liquid surface, large differences in temperature gradient at the solidification front, and a difference in billet shell thickness ≥ 0.5 mm. Microcracks have already appeared but have not propagated. The corresponding causes are improper nozzle immersion depth, excessively high casting speed, and excessively high superheat. The risk of severe longitudinal cracks indicates that when LFI≥0.6 and the model predicts a longitudinal crack probability≥95%, the flow velocity on the meniscus is severely uneven, the liquid level fluctuates significantly, the nozzle impact position shifts, the billet shell becomes locally thinner, microcracks have expanded, and obvious longitudinal cracks are easily formed, which may even lead to steel leakage. The corresponding causes are nozzle blockage, sudden change in casting speed, and abnormal crystallizer vibration parameters.

[0059] Example 1 like Figure 3 As shown, the present invention also includes a slab longitudinal crack prediction system based on the unbalanced flow field state of the crystallizer, comprising: a multi-source data acquisition module, a data preprocessing module, a longitudinal crack crystallizer flow field unbalance quantification module, a billet shell solidification dynamics constrained CNN-GRU longitudinal crack prediction module, and a longitudinal crack classification prediction module.

[0060] The multi-source data acquisition module is used to collect crystallizer flow field characteristic parameters, billet solidification characteristic parameters, and continuous casting process parameters that are strongly correlated with the formation of longitudinal cracks. It solves the problem of single-dimensional flow field and solidification data. It adopts online sensor acquisition and synchronous acquisition of process parameters to ensure data comprehensiveness and real-time performance, and meets the core requirements of longitudinal crack prediction.

[0061] The data preprocessing module is used to preprocess the raw collected data to remove noise interference, missing values, inconsistent dimensions, and spatiotemporal asynchrony, so as to ensure the accuracy of subsequent quantification and forecasting.

[0062] The longitudinal crack crystallizer flow field imbalance quantification module overcomes the problem of confusion between traditional flow field quantification and slag entrainment characteristics. Combining the formation mechanism of longitudinal crack shell inhomogeneity and stress concentration, it constructs a longitudinal crack-related flow field imbalance index (LFI) that integrates multi-dimensional flow field characteristics, thereby achieving accurate quantification of the flow field imbalance state and clarifying the correlation between flow field imbalance and longitudinal cracks.

[0063] The billet solidification dynamics constraint CNN-GRU longitudinal crack prediction module achieves high-precision longitudinal crack prediction by embedding billet solidification dynamics constraints and combining the spatial feature extraction capability of CNN with the temporal feature capture capability of GRU.

[0064] The longitudinal crack classification and prediction module combines the longitudinal crack-related flow field imbalance index (LFI) with the model prediction results and refers to the formation and propagation law of longitudinal cracks to classify the risk of longitudinal cracks into 3 levels, clarifying the classification basis, crack characteristics and inducing causes.

[0065] Example 2 This implementation method, based on a 200mm×1530mm low-carbon slab continuous casting production line of a steel plant's No. 6 continuous casting machine, details the optimal implementation method. This method covers equipment deployment, data acquisition, preprocessing, flow field imbalance quantification, model training, and online forecasting to ensure optimal forecasting results after implementation. Specific implementation details are as follows: The core equipment selection and system deployment utilize high-precision, interference-resistant acquisition and computing equipment adapted to the continuous casting industrial environment to build a dedicated longitudinal crack prediction system: The data acquisition end employs industrial-grade flow field feature acquisition equipment and solidification feature detection equipment to ensure accurate acquisition of core parameters such as meniscus velocity, liquid surface fluctuations, and solidification front temperature gradients; the computing end uses industrial computers to meet the real-time requirements of data preprocessing, LFI calculation, and model inference; the deployment end features a visual operation interface built in the continuous casting main control room, integrating data reception, LFI display, risk prediction, and early warning push; simultaneously, an industrial database is built to store multi-source data, model training data, and prediction and control records, supporting data backtracking and model optimization.

[0066] The calibration of core process and solidification parameters was conducted for 200mm×1530mm low-carbon slabs. Based on historical production data and experience in longitudinal crack control, reasonable ranges for core process parameters and solidification kinetic constraint thresholds for the slab shell were determined. These serve as the core basis for data verification, model constraints, and process control. Reasonable ranges for process parameters: slab casting speed 0.8-1.2m / min, submerged entry nozzle immersion depth 130-160mm, molten steel superheat 20-40℃, crystallizer vibration amplitude ±3mm, vibration frequency 120-220 times / min, and protective slag viscosity 0.12-0.28Pa. s; Solidification kinetics constraint thresholds: meniscus molten steel flow velocity ≤ 0.8 m / s, solidification front temperature gradient ≤ 8℃ / mm, crystallizer outlet billet shell critical thickness ≥ 3 mm (to avoid excessively thin billet shell cracking), liquid surface fluctuation amplitude ≤ ± 3 mm.

[0067] Dataset Construction and Preprocessing: A composite dataset combining numerical simulation and industrial field data was constructed to ensure model training covers multiple operating conditions and improve model generalization ability. A CFD solidification coupling model was built to simulate the interaction between molten steel flow and billet solidification within the crystallizer. The influence of process parameters such as casting speed, superheat, and nozzle depth on heat transfer and solidification rate was considered. Latin hypercube sampling was used to generate 1200 sets of flow-solidification correlation data within a reasonable range of process parameters, labeling core features such as meniscus velocity, temperature gradient, and reflow intensity. The longitudinal crack risk level (no risk, mild, moderate, severe) was labeled based on the solidification simulation results. Industrial Field Data Construction: 600 sets of full-condition data from the stable production phase of the production line were collected and processed synchronously. Data flow field characteristics, billet solidification characteristics, and process parameters were collected. By combining eddy current detection on the billet surface and manual re-inspection, the actual occurrence and risk level of longitudinal cracks corresponding to each data set were labeled (220 sets of no risk, 150 sets of mild cracks, 130 sets of moderate cracks, and 100 sets of severe cracks). Data set preprocessing: simulated data and field data were merged, invalid and abnormal data were removed, and the training set, validation set, and test set were divided in a 7:2:1 ratio. Each data set was stamped with a millisecond-level timestamp and labeled with the corresponding process parameters and solidification characteristics to ensure that the dataset can be directly used for model training and validation.

[0068] S1: Multi-source data acquisition We collect data from multiple sources and preprocess it to eliminate noise, fill in missing data, and unify dimensions, providing high-quality data input for subsequent flow field imbalance quantification and model prediction.

[0069] Flow field characteristic parameter acquisition (15Hz): Acquired the flow velocity of molten steel on both sides of the wide face meniscus, the amplitude and frequency of liquid surface fluctuation, the backflow intensity of molten steel on both sides of the narrow face, the impact position and impact intensity of molten steel at the nozzle, and the quantitative index of the uniformity of molten steel flow on the meniscus.

[0070] Solidification characteristic parameter acquisition of billet shell (10Hz): Acquire temperature gradient at different depths of solidification front on both sides of the wide face, real-time thickness distribution of billet shell, and solidification-melting cycle frequency of meniscus billet shell.

[0071] Continuous casting process parameter acquisition (1Hz): Acquire billet casting speed, sprue immersion depth, molten steel superheat, crystallizer vibration parameters, and protective slag viscosity.

[0072] Example operating conditions and data collected: casting speed 1.3 m / min, nozzle depth 140 mm, molten steel superheat 32 °C, protective slag viscosity 0.2 Pa. s; Left meniscus velocity 0.55 m / s, right meniscus 0.48 m / s; Liquid surface fluctuation amplitude 2.1 mm, fluctuation frequency 6 Hz; Narrow face left side backflow intensity 0.42 m / s, right side 0.38 m / s; Nozzle impact position deviation 3 mm; Wide face left side solidification front temperature gradient 4.2 ℃ / mm, right side 3.9 ℃ / mm; Average billet shell thickness 3.5 mm; Meniscus billet shell solidification-melting cycle frequency 2 times / min.

[0073] S2: Preprocessing stage The entire data preprocessing process involves targeted treatment of the raw collected data to eliminate interference, standardize data, and ensure that the data meets the requirements for subsequent quantization and model input. Noise Removal: For random high-frequency noise in parameters such as flow velocity and temperature gradient, a combination of 6-layer db5 wavelet transform decomposition and moving median averaging (moving window 12) is used for noise removal, controlling the noise of liquid surface fluctuation data within ±0.1mm and the noise of flow velocity data within ±0.003m / s; for visual data of flow morphology, Gaussian filtering is used for noise removal to preserve core flow features. Missing value imputation: Short-term missing values ​​(≤8 sampling points) are imputed using linear interpolation to ensure data continuity; long-term missing values ​​(>8 sampling points) are imputed using the weighted mean of historical data under the same working conditions to avoid the one-sidedness of imputation using a single mean. Spatiotemporal alignment: Based on the 10Hz solidification characteristic parameter timestamp of the billet shell, the 15Hz flow field characteristic parameter is downsampled to 10Hz (retaining characteristic peaks and not losing key information), and the 1Hz process parameter is frequency-complemented to 10Hz through linear interpolation. All data sampling frequencies are unified to 10Hz to ensure complete matching of flow field, solidification, and process data within the same time step. Outlier removal: using 3 The criteria combine process / solidification constraint thresholds for dual verification. Data exceeding the mean ± 3 times the standard deviation or violating the constraint threshold is judged as an outlier and replaced with the mean of normal data under the same operating conditions. Data normalization: min-max normalization is used to map all parameters to the [0,1] interval to eliminate dimensional differences. At the same time, the indicators such as meniscus velocity uniformity and temperature gradient uniformity, which are considered better the larger the value, are converted to the logic that the larger the value, the more unbalanced the flow field, ensuring that the contribution logic of all features to LFI is consistent.

[0074] S3: Quantification stage of flow field imbalance in longitudinal cracks The core flow field characteristics of longitudinal cracks were calculated based on preprocessed standardized data. Five core flow field characteristics that have the most significant impact on longitudinal cracks were calculated one by one: uniformity of molten steel velocity at the meniscus (F1). F1 = 1 - |left meniscus velocity - right meniscus velocity| / (left meniscus velocity + right meniscus velocity) = 1 - |0.55 - 0.48| / (0.55 + 0.48) = 0.932, which is converted to an imbalance eigenvalue of 1 - 0.932 = 0.068. After normalization... =0.068; Liquid surface fluctuation stability (F2): F2 = Standard deviation of liquid surface fluctuation amplitude / Mean of liquid surface fluctuation amplitude = 0.18, after normalization =0.18; Uniformity of temperature gradient at the solidification front (F3): F3 = 1 - |mean of temperature gradient on the left side - mean of temperature gradient on the right side| / (mean of temperature gradient on the left side + mean of temperature gradient on the right side) = 1 - |4.2 - 3.9| / (4.2 + 3.9) = 0.963, which is converted to an unevenness characteristic value of 1 - 0.963 = 0.037. After normalization... =0.037; Narrow face molten steel reflux intensity difference (F4): F4 = |left narrow face reflux intensity - right narrow face reflux intensity| = |0.42 - 0.38| = 0.04 m / s, after normalization =0.05; Impact position deviation of sprue (F5): F5 = |Actual impact position - Ideal impact position| / Ideal impact position = |3 - 0| / 10 = 0.3, after normalization =0.3; The combined weight matching and LFI calculation adopt a combination of the analytic hierarchy process (AHP), entropy weight method, and process correlation method for weight assignment. The final combined weight W = [0.277, 0.251, 0.22, 0.151, 0.101] is then substituted into the LFI calculation formula: LFI=0.277×0.068+0.251×0.18+0.22×0.037+0.151×0.05+0.101×0.3=0.0188+0.0452+0.0081+0.0076+0.0303=0.1099; The degree of imbalance is determined by combining the LFI classification threshold. If 0.1099 < 0.25, the current flow field is determined to be in equilibrium, and the basic risk level of longitudinal cracks is no risk.

[0075] S4: Training and Optimization Stage of the CNN-GRU Model for Solidification Kinetics Constraints on the Shell The model is built and trained, with solidification kinetic constraints embedded and hyperparameters optimized to ensure high accuracy and generalization ability, meeting the needs of industrial online real-time forecasting. Model deployment involves solidifying and deploying the trained and optimized model to the industrial computer in the main control room, enabling real-time inference mode to ensure that preprocessed multi-source data can be input into the model in real time and quickly output forecast results.

[0076] S5: Online Longitudinal Crack Grading Prediction and Closed-Loop Control Stage Under normal operating conditions, the pre-processed flow field, solidification, and process data are input into the model in real time. Combined with the LFI value, the model outputs the probability of longitudinal crack risk: no risk: 99.1%, slight risk: 0.7%, moderate risk: 0.1%, severe risk: 0.1%. The final prediction result is no risk of longitudinal crack. The main control room interface displays an LFI value of 0.1099, a flow field equilibrium state, and a no-risk conclusion. No warning is triggered, and production is maintained with the current process parameters.

[0077] Handling of Moderate Longitudinal Crack Risk Condition (Core Verification Scenario): Anomaly Trigger: Casting speed increased to 1.3 m / min (exceeding the reasonable range), nozzle depth decreased to 125 mm, LFI calculated after data preprocessing = 0.52, model output moderate longitudinal crack risk probability 92%, triggering an orange audible and visual alarm. Control Room Display: Moderate risk, core anomaly: excessive casting speed, violent liquid level fluctuations. Process Control: Operators executed the control plan according to instructions: casting speed decreased from 1.3 m / min to 1 m / min; nozzle depth increased from 125 mm to 140 mm; tundish insulation was adjusted, molten steel superheat decreased from 38℃ to 30℃. Closed-Loop Verification: 5 minutes after adjustment, data was collected and LFI calculated to 0.28, model predicted a slight risk probability of 65%; casting speed was further fine-tuned to 1.2 m / min, LFI = 0.22 after 1 minute, model predicted no risk, warning was lifted, flow field returned to equilibrium, preventing longitudinal crack formation.

[0078] An anomaly was triggered in the emergency response to a severe longitudinal crack risk condition: a slight blockage at the nozzle caused a significant shift in the impact position, raising the LFI to 0.75. The model output a 98% probability of a severe longitudinal crack, triggering a red emergency alarm. The main control room then pushed out an emergency control plan. Emergency control measures: immediately reduce the casting speed to 0.8 m / min to extend the solidification time of the billet shell; adjust the nozzle depth to 165 mm (upper limit) to stabilize the molten steel flow trajectory; suspend argon supply to eliminate surface disturbance; check that the nozzle is not blocked and adjust the crystallizer vibration frequency to 180 times / min; closed-loop verification: after 10 minutes, the LFI dropped to 0.35, and the risk decreased to mild; continue adjusting the casting speed to 1.1 m / min, and finally, the LFI = 0.21, indicating no risk was predicted, the warning was lifted, and crack propagation and steel leakage accidents were avoided.

[0079] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer, characterized in that, include: Collect crystallizer flow field characteristic parameters, billet shell solidification characteristic parameters, and continuous casting process parameters, and construct an operating parameter dataset; The dataset of operating parameters is preprocessed and normalized. Filter the flow field features that lead to longitudinal cracks from the aforementioned operating parameter dataset; The flow field characteristics leading to longitudinal cracks include: uniformity of molten steel flow velocity at the meniscus, stability of liquid surface fluctuations, uniformity of temperature gradient at the solidification front, difference in backflow intensity of molten steel at the narrow face, and offset of nozzle impact position, wherein: The uniformity of the molten steel flow velocity on the meniscus The formula used to represent the difference in flow velocity on both sides of the meniscus is: in, , The flow velocity on both sides of the meniscus is the flow velocity. The value range is 0-1; Liquid level fluctuation stability The coefficient of variation, used to represent the amplitude of liquid level fluctuations in the crystallizer, is calculated using the following formula: in, Let A be the standard deviation of the liquid level fluctuation amplitude, and let A be the mean of the liquid level fluctuation amplitude. The value is greater than 0; The uniformity of the temperature gradient at the solidification front The formula used to represent the difference in temperature gradient at the solidification front at the same depth on both sides of the wide face is: in, , The average temperature gradient on both sides of the wide surface is represented by F3, which ranges from 0 to 1. The narrow-faced molten steel reflow intensity difference The formula used to represent the difference in the intensity of molten steel reflow on both narrow sides is as follows: in, , For the backflow intensity on both narrow sides, the value of F4 is greater than 0; The offset of the water inlet impact position The formula used to indicate the deviation between the actual impact position and the ideal impact position of molten steel at the nozzle is as follows: Where L represents the actual impact location. For the ideal impact location, the value of F5 is greater than 0; The flow field characteristics leading to longitudinal cracks were weighted using the analytic hierarchy process (AHP), entropy weighting method, and process correlation method, and combined weights were obtained. The longitudinal crack-related flow field imbalance index is calculated based on the combined weights, and the degree of longitudinal crack-related flow field imbalance is classified. A longitudinal crack prediction model is constructed, and the longitudinal crack prediction model is trained and optimized. The longitudinal crack is predicted in stages by combining the longitudinal crack-related flow field imbalance index and the output results of the longitudinal crack prediction model.

2. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 1, characterized in that, The flow field characteristic parameters of the crystallizer include: meniscus molten steel flow velocity, crystallizer liquid surface fluctuation characteristics, narrow-face molten steel backflow intensity, and molten steel impact characteristics at the nozzle; the solidification characteristic parameters of the billet shell include: solidification front temperature gradient, billet shell thickness distribution characteristics, and meniscus billet shell solidification-melting cycle frequency; the continuous casting process parameters include: billet casting speed, submerged nozzle immersion depth, molten steel superheat, crystallizer vibration parameters, and protective slag viscosity.

3. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 1, characterized in that, The data preprocessing includes: noise removal, missing value imputation, spatiotemporal alignment, and outlier removal; The noise removal method refers to using a combination of wavelet transform and moving median averaging for random noise removal to eliminate high-frequency interference and burst noise; and using Gaussian filtering for high-speed camera image noise removal to preserve meniscus flow edge features. The missing value imputation refers to the imputation of missing values ​​caused by temporary sensor failures. Short-term missing values ​​are imputed by linear interpolation, while long-term missing values ​​are imputed by the weighted average of historical data under the same operating conditions to ensure data continuity. The spatiotemporal alignment means that, based on the 10Hz solidification front temperature parameter timestamp, the flow field characteristic parameters are downsampled to 10Hz, and the process parameters are linearly interpolated to 10Hz, with a unified sampling frequency of 10Hz, to ensure data integrity within the same time step. The outlier removal means using... The criteria combine process boundary constraints to eliminate outlier data. When data exceeds the mean standard deviation or the reasonable range of the process, it is judged as an outlier and replaced with the mean of the same operating conditions. The normalization process maps all feature parameters to the [0,1] interval, eliminates dimensional differences, and uses the min-max normalization formula: in, The original data, This represents the normalized data; , These represent the minimum and maximum values ​​of this parameter under the same operating conditions.

4. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 1, characterized in that, The analytic hierarchy process (AHP) is based on continuous casting technology and the mechanism of longitudinal cracking. It constructs a judgment matrix to determine the subjective weights of the flow field characteristics that lead to longitudinal cracks. The entropy weight method, based on historical production data, calculates the information entropy of each feature to determine the objective weights of the flow field features that lead to longitudinal cracks. The process correlation method determines the correlation weight of the flow field characteristics that lead to the longitudinal crack by calculating the correlation between each feature and the longitudinal crack. ; The combined weights Represented as: in, These are the weighting coefficients corresponding to the subjective weights. This refers to the weighting coefficient corresponding to the objective weight.

5. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 4, characterized in that, The longitudinal crack-related flow field imbalance index (LFI) is calculated based on the combined weights and normalized eigenvalues, using the following formula: in, For the normalized first 1 eigenvalue, For the first The combined weights corresponding to the flow field characteristics that lead to longitudinal cracks; Based on the value of the longitudinal crack-associated flow field imbalance index (LFI) and historical longitudinal crack occurrence data, the degree of flow field imbalance is divided into three categories: balanced, unbalanced, and severely unbalanced. When 0 ≤ LFI < 0.25, the risk of longitudinal cracks is low, which is the balanced state. When 0.25 ≤ LFI < 0.6, the risk of longitudinal cracks is high, which is the unbalanced state. When LFI ≥ 0.6, the risk of longitudinal cracks is extremely high, which is the severely unbalanced state.

6. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 1, characterized in that, The longitudinal crack prediction model outputs longitudinal crack prediction results by embedding solidification dynamics constraints and process solidification coupling constraints of billet shell, and combining the spatial feature extraction capability of CNN and the temporal feature capture capability of GRU. The embedded billet shell solidification dynamics constraint is based on the solidification dynamics principle of continuous casting billet shell, and adds the following constraints: meniscus molten steel flow velocity M≤0.8m / s; solidification front temperature gradient N≤8℃ / mm; billet shell critical thickness O≥3mm. When the input data violates the embedded billet shell solidification dynamics constraint, it is directly marked as high risk and a warning is triggered first. The process solidification coupling constraint embeds the correlation between process parameters and solidification characteristics into the longitudinal crack prediction model. Increasing the casting speed will lead to a shorter solidification time for the billet shell, and higher superheat will reduce the strength of the billet shell. This guides the model to learn the coupling relationship and improves its adaptability across working conditions.

7. The method for predicting longitudinal cracks in slabs based on the unbalanced flow field in the crystallizer according to claim 1, characterized in that, The graded forecast combines the longitudinal crack-related flow field imbalance index (LFI) with the output of the longitudinal crack forecasting model, and refers to the formation and propagation laws of longitudinal cracks to classify the longitudinal crack risk into mild longitudinal crack risk, moderate longitudinal crack risk and severe longitudinal crack risk. The risk of mild longitudinal cracks is defined as follows: when 0.25≤LFI<0.4, the model predicts a longitudinal crack probability of 60%-80%, the flow velocity on the meniscus is slightly uneven, the liquid level fluctuates slightly, the thickness of the billet shell is slightly different, there are no obvious cracks, and there is only a potential risk of cracking; the corresponding inducing factors are slight fluctuations in the casting speed and unstable argon flow rate. The moderate longitudinal crack risk refers to a situation where 0.4 ≤ LFI < 0.6, the model predicts a longitudinal crack probability of 80%-95%, the meniscus flow velocity is significantly uneven, the liquid surface fluctuates violently, the temperature gradient at the solidification front is large, the shell thickness difference is ≥ 0.5 mm, and microcracks have appeared but have not propagated. The corresponding causes are improper nozzle immersion depth, excessively high casting speed, and excessively high superheat. The severe longitudinal crack risk indicates that when LFI≥0.6 and the model predicts a longitudinal crack probability≥95%, the meniscus flow velocity is severely uneven, the liquid level fluctuates significantly, the nozzle impact position shifts, the billet shell becomes locally thinner, microcracks have expanded, and obvious longitudinal cracks are easily formed, which may even lead to steel leakage. The corresponding inducing causes are nozzle blockage, sudden changes in casting speed, and abnormal crystallizer vibration parameters.

Images