High-titanium steel fluoride-free continuous casting protective slag with high heat transfer control ability and preparation method thereof

Abstract

The application discloses a high-titanium steel fluoride-free continuous casting protective slag with strong heat transfer control capability and a preparation method thereof. In the application, an online optimization module based on Gaussian process regression is introduced in a key water quenching and rapid cooling link. By collecting working condition parameters such as melt temperature, cooling water temperature and environmental humidity in real time, the optimal pouring duration and liquid flow impact angle are automatically recommended by a data-driven model, the glass transition rate of slag blocks after water quenching is stably controlled at a high level, the uniformity of each batch of raw materials in the organizational structure is ensured, and therefore, the microstructure level guarantee for the consistency of the crystallization behavior of the subsequent solid slag film and the heat transfer stability is provided. The traditional manual water quenching operation depending on the experience of operators is upgraded to a closed-loop control system with self-learning capability, and the batch consistency of the product quality is significantly improved without increasing equipment investment.

Description

Technical Field

[0001] This invention belongs to the field of high-titanium steel technology, specifically a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability and its preparation method. Background Technology

[0002] High-titanium steel, due to its high strength, good corrosion resistance, and excellent low-temperature toughness, has been widely used in aerospace, automotive manufacturing, petrochemical, and medical device industries. With the continuous improvement of performance requirements for steel materials across various industries, the production volume and continuous casting demand for high-titanium steel are steadily increasing. However, a series of unique technical challenges exist in the continuous casting process of high-titanium steel, severely restricting its large-scale production and improving billet quality. The core issue in the continuous casting of high-titanium steel lies in the interfacial reaction between titanium in the molten steel and the components of the protective slag. During the casting process, titanium in the steel readily undergoes a violent interfacial reaction with silica in the protective slag, generating high-melting-point compounds such as titanium dioxide. This reaction not only leads to uncontrollable deviations from the original composition of the protective slag and severe deterioration of the slag's physicochemical properties, but also causes defects such as blockage of the submerged entry nozzle, slag formation on the surface of the molten steel in the crystallizer, and cracks, depressions, and inclusions on the billet surface, even resulting in steel leakage accidents in severe cases. Traditional protective slag systems based on calcium oxide and silica are insufficient to fundamentally eliminate the performance failures caused by the interfacial reaction. Therefore, the development of special protective slags for high-titanium steel with low reactivity or even non-reactivity has become a research hotspot and major trend in this field.

[0003] However, the volatilization of fluorides in existing fluorine-containing protective slags causes equipment corrosion and environmental pollution. High silica content can easily trigger steel-slag interface reactions, leading to deterioration of slag film performance. At the same time, key preparation processes rely on manual experience and judgment, resulting in significant fluctuations in the glass transition rate of different batches of slag blocks. The heat flow stability of the crystallizer is poor during continuous casting, and the risk of surface cracks and steel leakage in the billet remains high. Summary of the Invention

[0004] The purpose of this invention is to provide a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability and its preparation method in order to solve the problems mentioned above.

[0005] The technical solution adopted in this invention is as follows: a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability, comprising: CaO: 28 parts by weight, Al2O3... 3: 26 parts by weight, TiO 2: 8 parts by weight, Na₂O: 10 parts by weight, SiO 2: 5 parts by weight, BaO: 12 parts by weight, B2O 3: 11 parts by weight.

[0006] In a preferred embodiment, a method for preparing a fluorine-free continuous casting protective slag for high-titanium steel with strong heat transfer control capability is characterized by comprising the following steps: S1: Calculate and weigh the chemically pure raw materials CaO, Al2O3, TiO2, SiO2, B2O3, Na2CO3 and BaCO3 according to the formula ratio; before weighing, dry all raw materials in an oven at 105~110℃ for 2~4 hours to remove free moisture.

[0007] S2: Load all the weighed materials into the planetary ball mill jar, control the ball-to-material ratio to be 1, and dry ball mill for 30 minutes to make each component reach a fully uniform powder state.

[0008] S3: Transfer the mixed powder to a high-purity graphite crucible, place it in a medium-frequency induction furnace, and heat it to 1400℃ at a heating rate of 5~8℃ / min. After the material is completely melted, keep it at that temperature for 20 minutes. During the holding period, stir it gently with a corundum rod every 5 minutes to ensure that the melt is uniform.

[0009] S4: Immediately after the heat preservation is completed, the crucible is removed. The optimal pouring parameters are recommended based on the real-time data from the sensors using the Gaussian process regression model. The operator then continuously pours the high-temperature melt into a room-temperature deionized water bath for rapid cooling and quenching to obtain glassy slag. After water quenching, the glass transition rate and corresponding parameters are fed back to the model through rapid XRD detection to achieve iterative optimization for subsequent batches.

[0010] S5: Remove the water-quenched slag blocks, spread them on a stainless steel tray, and dry them in a 105℃ forced-air drying oven for 4-6 hours to remove the attached water.

[0011] S6: The dried slag is first coarsely crushed to a particle size of less than 5mm using a jaw crusher, and then transferred to a disc crusher or vibrating mill for fine grinding to further refine the particles.

[0012] S7: Pass the finely ground powder through a 200-mesh standard sieve, take the sieve material and place it in a V-type mixer to continue mixing for 15 minutes to eliminate local segregation that may be caused by grinding, and then seal and package it to obtain the finished product.

[0013] In a preferred embodiment, in step S1, all raw materials must be placed separately in a forced-air drying oven before weighing and continuously heated and dried in a temperature range of 105 ℃ to 110 ℃ for 2 h to 4 h to completely remove the free moisture adsorbed on the surface and inside of the particles, and prevent melt splashing or composition fluctuations caused by moisture evaporation during high-temperature melting. After drying, each raw material is taken out and naturally cooled to room temperature for later use.

[0014] In a preferred embodiment, in step S2, zirconia balls are added in a ratio of 1:2 to 1:3 between the total mass of the powder and the mass of the grinding media. After the can lid is tightened, the ball mill is started and dry mixing and grinding are performed at a spindle speed of 250 r / min to 300 r / min for a total grinding time of 30 min. During the grinding process, the mill is paused every 10 min to prevent the can from overheating. After the grinding is completed, the mill is stopped and left to stand for 5 min to allow the dust in the can to settle naturally. Then the can lid is opened and the mixed powder is taken out. At this time, the powder should be uniform in color and free from visible agglomeration or stratification.

[0015] In a preferred embodiment, in step S3, after closing the furnace door, the heating program is started, and the heating rate is controlled to be maintained between 5 ℃ / min and 8 ℃ / min. When the temperature reaches 1400 ℃, the heat preservation mode is switched and the temperature is maintained for 20 min. During the heat preservation period, every 5 min, a corundum rod with a diameter of about 6 mm is vertically inserted into the crucible molten pool from the top of the furnace and gently stirred for one revolution. Each stirring lasts for about 30 seconds. After stirring, the corundum rod is immediately pulled out and the feeding port is covered. This operation is repeated 4 times to ensure that the temperature field and concentration field in the molten pool are uniformly distributed.

[0016] In a preferred embodiment, in step S4, a special crucible tongs are used to hold the edge of the graphite crucible and quickly remove it from the furnace. The crucible is then transferred to a pre-prepared stainless steel water tank. The tank contains deionized water at room temperature, and the amount of water should be sufficient to completely submerge the entire melt. The operator slowly tilts the crucible so that the high-temperature melt flows into the water in a continuous, thin stream. The entire tilting process is completed within 5 to 8 seconds. The melt undergoes a rapid cooling and quenching effect upon entering the water, resulting in transparent or translucent glassy slag particles. After water quenching, the slag is left to cool in the water until the water temperature returns to room temperature.

[0017] In a preferred embodiment, in step S4, the crucible is immediately removed and transferred to the water quenching station after the heat preservation is completed. At this time, the online optimization module based on Gaussian process regression is started. This module needs to pre-construct an initial training set, which can be sourced from previous orthogonal experimental data or historical records of similar production lines. Each sample in the training set consists of an input feature vector and an output label. The input features include five key parameters during the water quenching process: the actual temperature of the melt, the pouring duration, the temperature of the deionized water, the vertical distance from the crucible opening to the liquid surface, and the relative humidity of the environment. The output label is the glass transition rate of the slag. The glass transition rate is obtained by randomly selecting no less than three slag samples from the batch, grinding and pressing them into tablets, and then performing a full XRD spectrum scan. The ratio of the area of ​​the amorphous scattering peaks to the total scattering peak area is fitted by phase analysis software. The initial training set contains at least 20 sets of valid data pairs to ensure prior reliability. This prediction step is achieved by the posterior mean formula and the posterior variance formula. The former outputs the predicted value, and the latter quantifies the confidence interval. The two constitute the basis for model inference.

[0018] The core of Gaussian process regression lies in assigning a priori values ​​to the objective function, determined by the mean and covariance functions. This module selects the Matern 5 / 2 kernel function to handle the nonlinear relationship between process parameters and glass transition rate during water quenching. This kernel function can maintain reasonable prediction uncertainty even with limited training data. After the initial training set is established, hyperparameters such as signal variance, length scale, and noise variance are optimized using the logarithmic marginal likelihood function to complete model training. When the S3 step of heat preservation ends, the infrared temperature sensor placed at the induction furnace outlet automatically transmits the real-time melt temperature to the model, while the temperature and humidity sensors at the water quenching station simultaneously input cooling water temperature and ambient humidity data. The model uses these three measurable parameters and the two operational parameters to be optimized to form the query point, and utilizes the Gaussian process combined with the normal distribution property to quickly calculate the predicted mean and variance of glass transition rate.

[0019] The Bayesian optimization framework replaces manual experience-based decision-making here, automatically recommending the optimal water quenching parameters. The optimization objective is to maximize the glass transition rate of the slag while constraining the pouring duration and liquid flow impact angle to remain within the safe range allowed by the equipment. This module uses the expected improvement formula as the acquisition function, which balances the potential for improvement of the predicted value relative to the current optimal value with the uncertainty of the prediction, automatically balancing the contradiction between development and exploration. The acquisition function performs global maximization in the two-dimensional parameter space through grid partitioning and differential evolution algorithm, completing the calculation in milliseconds and displaying the optimal parameters on the industrial display screen. Based on this, the operator pours the high-temperature melt in a continuous thin stream into a room-temperature deionized water bath for rapid cooling, resulting in glassy slag fragments.

[0020] After water quenching, the actual effect needs to be fed back to the model to support iterative optimization. Randomly selected batches of slag are subjected to rapid XRD detection. The measured glass transition rate and the current process parameters are used to form new sample points, which are added to the training set incrementally. Every 10 newly added samples trigger an incremental update, i.e., the logarithmic marginal likelihood maximization is re-executed using the previously optimized hyperparameters to achieve local fine-tuning. Simultaneously, the covariance matrix incremental update formula is used to accelerate subsequent prediction calculations. As batches accumulate, the model's posterior distribution gradually approximates the true mapping relationship, and the recommendation accuracy continuously improves, ultimately forming a data-driven water quenching closed-loop control system with self-learning capabilities. The expression for the predicted posterior mean at the query point using Gaussian process regression is: in The input feature vector at the query point represents the melt temperature, water temperature, ambient humidity, and two operational parameters to be optimized, all collected by the current sensor. X represents the matrix composed of all input samples in the training set. y represents the output label vector of the training set, i.e., the glass transition rate corresponding to each batch. k( X) represents the covariance vector between the query point and each sample in the training set, which is calculated by the kernel function; K is the covariance matrix between the samples in the training set. 2 n is the noise variance hyperparameter; I is the identity matrix. This formula outputs the predicted average glass transition rate for the query point, given the current training data.

[0021] The expression for the predicted posterior variance is: Where k( , ) represents the covariance between the query point and itself, i.e., the maximum value of the kernel function; k(X, ) and k( X and X are transposes of each other. This formula gives the magnitude of the uncertainty in the predicted value, and its magnitude directly determines the strength of the exploration tendency in the subsequent acquisition function. For the Matern5 / 2 kernel function, its specific form is: , in This represents the Euclidean distance between two input points. 2 f is the signal variance hyperparameter, and l is the length scale hyperparameter. These two hyperparameters, together with the noise variance, constitute the set of kernel hyperparameters to be optimized.

[0022] The log-marginal likelihood function used during the model training phase is: Wherein represents the set of hyperparameters, including signal variance, length scale, and noise variance; K represents the covariance matrix of the training set samples given the hyperparameters; and n is the total number of training set samples. This function represents matrix determinant operations. The first term on the right-hand side measures how well the model fits the training data, the second term acts as a penalty to control model complexity and prevent overfitting, and the third term is a constant. Maximizing this function yields the optimal hyperparameter estimates.

[0023] The desired improvement acquisition function used in Bayesian optimization is: Where f+ represents the maximum glass transition rate observed in the current training set; (x) and (x) are the predicted mean and predicted standard deviation at the query point, respectively, given by the first two equations; Let represent the cumulative distribution function and probability density function of the standard normal distribution, respectively. The first term on the right-hand side of the equation encourages development behavior when the predicted mean is higher than the current optimal value, and the second term encourages exploration behavior when the predicted standard deviation is large. The input point corresponding to the global maximum value of this function is the optimal operating parameter recommended for the next batch.

[0024] The incremental update expression for the covariance matrix is ​​used to support efficient iteration of the model as data accumulates: Where the subscript t represents the current training set size, x t+1 For the newly added sample points, K t The old covariance matrix, This represents the covariance vector between the new sample point and each sample in the old training set. This incremental update avoids the computational burden of recalculating the entire covariance matrix after each new sample, enabling the model to maintain real-time responsiveness in large-scale continuous production scenarios. All formulas in this step are executed simultaneously, collectively realizing the transformation of the water quenching process from manual experience-based operation to data-driven intelligent decision-making.

[0025] In a preferred embodiment, in step S5, the water-quenched slag particles are scooped out of the water tank with a stainless steel strainer, drained of a large amount of free water, and then evenly spread on a stainless steel tray. The spreading thickness should not exceed 3 cm to ensure uniformity of subsequent drying. The tray is placed in a forced-air drying oven and the temperature is set to 105 degrees Celsius. Drying is carried out for 4 to 6 hours. During the drying period, the slag particles are turned over every 2 hours to accelerate moisture evaporation and prevent the particles from sticking together. The drying endpoint is judged by the mass change not exceeding 0.1% within a 30-minute weighing interval. After drying is completed, the oven power is turned off, and the slag blocks are allowed to cool naturally in the oven to a temperature that is not hot to the touch before being taken out.

[0026] In a preferred embodiment, in step S6, the discharge gap of the crusher is adjusted so that the maximum size of the crushed particles is controlled below 5 mm. After coarse crushing, the obtained particles are transferred to a disc crusher for medium crushing and fine grinding. The disc spacing is gradually tightened so that the material undergoes splitting, grinding and shearing in sequence. Finally, the discharge particle size reaches the level where all particles pass through a 150-mesh standard sieve. If the grinding disc is found to be significantly hot during the fine grinding process, it is paused and allowed to cool naturally before continuing to prevent local high temperature from causing phase change of the powder or adhesion to the working surface of the grinding disc.

[0027] In a preferred embodiment, in step S7, the mixer speed is set to 12 to 15 revolutions per minute, and the mixture is continuously mixed for 15 minutes to eliminate local segregation that may be caused by density differences during the previous crushing and grinding process. After mixing, the machine is stopped and the discharge port is opened. The powder is then passed through a 200-mesh standard vibrating screen for final sieving. The undersize material is taken as the qualified finished product, and the oversize material is returned to the disc pulverizer for fine grinding and sieving again. The qualified powder is placed in a double-layer moisture-proof plastic bag, the air inside the bag is removed, and the bag is heat-sealed. It is then placed in a sealed metal barrel for storage in the dark and dry place. A label is attached to the outside of the barrel indicating the sample number, production date, and sieving particle size information. This completes all the preparation steps of the high-titanium steel fluorine-free continuous casting protective slag.

[0028] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. In this invention, titanium dioxide completely replaces fluoride, which induces the preferential precipitation of perovskite phase in the slag film, resulting in a significant increase in the thickness of the solid slag film and a substantial improvement in surface roughness and porosity. At the same time, silicon dioxide is strictly controlled at a low level to suppress the steel-slag interface reaction during the casting of high-titanium steel, thereby effectively suppressing the heat flux density in the later stage of solidification. This solves the problem of unstable heat transfer caused by the different precipitation behavior of gun crystals in traditional fluorine-containing protective slags, while avoiding the corrosion of equipment and environmental pollution caused by fluoride, thus taking into account both green manufacturing and excellent lubrication performance.

[0029] 2. In this invention, an online optimization module based on Gaussian process regression is introduced in the critical water quenching and rapid cooling process. By collecting operating parameters such as melt temperature, cooling water temperature, and ambient humidity in real time, the data-driven model automatically recommends the optimal pouring duration and liquid flow impact angle, which stably controls the glass transition rate of the slag block after water quenching at a high level, ensuring the uniformity of the microstructure of each batch of raw materials. This provides a microstructural guarantee for the consistency of the crystallization behavior and heat transfer stability of the subsequent solid slag film. This method upgrades the traditional manual water quenching operation that relies on operator experience to a closed-loop control system with self-learning capabilities, significantly improving the batch consistency of product quality without increasing equipment investment. Attached Figure Description

[0030] Figure 1This is a schematic diagram illustrating the process principle of the present invention; Figure 2 This is a comparison diagram of the slag film structure and heat transfer performance of different protective slags in this invention; Figure 3 This is a comparison chart showing the effects of different protective slags on continuous casting in this invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] Reference Figure 1-3 , Example 1 This embodiment provides a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability, comprising, by weight: 28 parts CaO, 26 parts Al2O3, 28 parts TiO, 10 parts Na2O, 5 parts SiO, 12 parts BaO, and 11 parts B2O3, without the addition of CaF2. Its preparation method is as follows: S1: Place the chemically pure raw materials CaO, Al2O3, TiO2, SiO2, B2O3, and the precursors Na2CO3 and BaCO3 into a forced-air drying oven and dry them at 105 ℃ to 110 ℃ for 3 h to remove free moisture. Then, calculate and weigh each raw material according to the formula.

[0033] S2: Load all the weighed powder into the grinding jar of the planetary ball mill at once, control the ball-to-powder ratio to be 2.5:1, and dry mix at a speed of 280 r / min for 30 min.

[0034] S3: Transfer the uniformly mixed powder to a high-purity graphite crucible, place it in a medium-frequency induction furnace and heat it to 1400 ℃ at a heating rate of 6 ℃ / min. After the material is completely melted, keep it at this temperature for 20 min. During the holding period, stir it once every 5 min with an alumina rod for about 30 seconds each time.

[0035] S4: After the heat preservation is completed, the crucible is immediately removed. The operator judges the pouring speed and angle based on personal experience and pours the high-temperature melt into the room temperature deionized water bath at a manually controlled rate to complete the water quenching and rapid cooling, and obtains glassy slag blocks.

[0036] S5: After removing the water-quenched slag, spread it evenly on a stainless steel tray and dry it in a 105 ℃ forced-air drying oven for 5 hours.

[0037] S6: The dried slag is coarsely crushed to a particle size of less than 5 mm by a jaw crusher, and then transferred to a disc mill for fine grinding until all of it passes through a 150-mesh standard sieve.

[0038] S7: Pass the finely ground powder through a 200-mesh vibrating screen for final sieving. Take the undersize material and mix it in a V-type mixer at 15 r / min for 15 min. Seal and package to obtain the finished product.

[0039] The only difference between this embodiment and embodiment 2 is that the online optimization module of machine learning was not introduced in the S4 water quenching process. The pouring duration and liquid flow impact angle are determined by the operator based on on-site experience.

[0040] Example 2 This embodiment provides a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability. Its formula is exactly the same as in Example 1: 28 parts CaO, 26 parts Al2O3, 28 parts TiO, 10 parts Na2O, 5 parts SiO, 12 parts BaO, and 11 parts B2O3, without the addition of CaF2. Steps S1, S2, S3, S5, S6, and S7 in its preparation method are completely identical to those in Example 1, except that the online optimization module based on Gaussian process regression is activated in the S4 water quenching step, as detailed below: S1: Same as in Example 1.

[0041] S2: Same as Example 1.

[0042] S3: Same as Example 1.

[0043] S4: Immediately after the heat preservation is completed, remove the crucible and transfer it to the water quenching station, and start the online optimization module based on Gaussian process regression.

[0044] This module pre-constructs an initial training set using 28 sets of prior orthogonal experimental data. Input features include five key process parameters: actual melt temperature, pouring duration, deionized water temperature, vertical distance from the crucible opening to the liquid surface, and ambient relative humidity. The output label is the glass transition rate of the corresponding batch of slag after water quenching. The method for determining the glass transition rate is as follows: At least three slag samples were randomly selected from the batch, ground, and pressed into tablets. X-ray diffraction full-spectrum scanning was then performed. The diffraction peak areas of each phase were fitted using phase analysis software, and the proportion of the amorphous scattering peak area to the total scattering peak area was defined as the glass transition rate. The model trained the Matern 5 / 2 kernel hyperparameters using the logarithmic marginal likelihood function. The glass transition rate was predicted using the posterior mean and posterior variance formulas based on the real-time operating conditions transmitted from the sensor and the operating parameters to be optimized. Global optimization was performed in the two-dimensional operating space using the desired improvement formula as the acquisition function.

[0045] When the S3 heat preservation step ends, the infrared temperature sensor located at the induction furnace outlet measures the actual melt temperature as 1412 ℃ and automatically transmits it to the model. Simultaneously, the temperature sensor at the water quenching station and the ambient humidity sensor input the deionized water temperature of 24.5 ℃ and the ambient relative humidity of 62% into the model. After millisecond-level calculations, the model recommends on the industrial display screen the optimal pouring duration for this batch as 6.2 s and the liquid flow impact angle as 58°.

[0046] Based on these parameters, the operator performs a water quenching and rapid cooling operation, pouring the high-temperature melt into a room-temperature deionized water bath in a continuous, thin stream. The resulting slag fragments, after rapid XRD analysis, show a glass transition rate of 94.3%. This set of measured data, combined with the corresponding process parameters, constitutes new sample points, which are added to the training set using an incremental update formula for the covariance matrix. Every 10 newly added samples trigger a local fine-tuning of hyperparameters for subsequent batch iterations and optimizations.

[0047] In step S4, the crucible is immediately removed and transferred to the water quenching station after the heat preservation is completed. At this time, the online optimization module based on Gaussian process regression is started. This module requires the pre-construction of an initial training set, which can be sourced from previous orthogonal experimental data or historical records of similar production lines. Each sample in the training set consists of an input feature vector and an output label. The input features include five key parameters during the water quenching process: the actual melt temperature, the pouring duration, the deionized water temperature, the vertical distance from the crucible opening to the liquid surface, and the ambient relative humidity. The output label is the glass transition rate of the slag. The glass transition rate is obtained by randomly selecting no less than three slag samples from the batch, grinding and pressing them into tablets, and then performing a full XRD spectrum scan. The ratio of the amorphous scattering peak area to the total scattering peak area is fitted by phase analysis software. The initial training set contains at least 20 sets of valid data pairs to ensure prior reliability. This prediction step is achieved by the posterior mean formula and the posterior variance formula. The former outputs the predicted value, and the latter quantifies the confidence interval. The two constitute the basis for model inference.

[0048] The core of Gaussian process regression lies in assigning a priori values ​​to the objective function, determined by the mean and covariance functions. This module selects the Matern 5 / 2 kernel function to handle the nonlinear relationship between process parameters and glass transition rate during water quenching. This kernel function can maintain reasonable prediction uncertainty even with limited training data. After the initial training set is established, hyperparameters such as signal variance, length scale, and noise variance are optimized using the logarithmic marginal likelihood function to complete model training. When the S3 step of heat preservation ends, the infrared temperature sensor placed at the induction furnace outlet automatically transmits the real-time melt temperature to the model, while the temperature and humidity sensors at the water quenching station simultaneously input cooling water temperature and ambient humidity data. The model uses these three measurable parameters and the two operational parameters to be optimized to form the query point, and utilizes the Gaussian process combined with the normal distribution property to quickly calculate the predicted mean and variance of glass transition rate.

[0049] The Bayesian optimization framework replaces manual experience-based decision-making here, automatically recommending the optimal water quenching parameters. The optimization objective is to maximize the glass transition rate of the slag while constraining the pouring duration and liquid flow impact angle to remain within the safe range allowed by the equipment. This module uses the expected improvement formula as the acquisition function, which balances the potential for improvement of the predicted value relative to the current optimal value with the uncertainty of the prediction, automatically balancing the contradiction between development and exploration. The acquisition function performs global maximization in the two-dimensional parameter space through grid partitioning and differential evolution algorithm, completing the calculation in milliseconds and displaying the optimal parameters on the industrial display screen. Based on this, the operator pours the high-temperature melt in a continuous thin stream into a room-temperature deionized water bath for rapid cooling, resulting in glassy slag fragments.

[0050] After water quenching, the actual effect needs to be fed back to the model to support iterative optimization. Randomly selected batches of slag are subjected to rapid XRD detection. The measured glass transition rate and the current process parameters are used to form new sample points, which are added to the training set incrementally. Every 10 newly added samples trigger an incremental update, i.e., the logarithmic marginal likelihood maximization is re-executed using the previously optimized hyperparameters to achieve local fine-tuning. Simultaneously, the covariance matrix incremental update formula is used to accelerate subsequent prediction calculations. As batches accumulate, the model's posterior distribution gradually approximates the true mapping relationship, and the recommendation accuracy continuously improves, ultimately forming a data-driven water quenching closed-loop control system with self-learning capabilities. The expression for the predicted posterior mean at the query point using Gaussian process regression is: in The input feature vector at the query point represents the melt temperature, water temperature, ambient humidity, and two operational parameters to be optimized, all collected by the current sensor. X represents the matrix composed of all input samples in the training set. y represents the output label vector of the training set, i.e., the glass transition rate corresponding to each batch. k( X) represents the covariance vector between the query point and each sample in the training set, which is calculated by the kernel function; K is the covariance matrix between the samples in the training set. 2 n is the noise variance hyperparameter; I is the identity matrix. This formula outputs the predicted average glass transition rate for the query point, given the current training data.

[0051] The expression for the predicted posterior variance is: Where k( , ) represents the covariance between the query point and itself, i.e., the maximum value of the kernel function; k(X, ) and k( X and X are transposes of each other. This formula gives the magnitude of the uncertainty in the predicted value, and its magnitude directly determines the strength of the exploration tendency in the subsequent acquisition function. For the Matern5 / 2 kernel function, its specific form is: , in This represents the Euclidean distance between two input points. 2 f is the signal variance hyperparameter, and l is the length scale hyperparameter. These two hyperparameters, together with the noise variance, constitute the set of kernel hyperparameters to be optimized.

[0052] The log-marginal likelihood function used during the model training phase is: Wherein represents the set of hyperparameters, including signal variance, length scale, and noise variance; K represents the covariance matrix of the training set samples given the hyperparameters; and n is the total number of training set samples. This function represents matrix determinant operations. The first term on the right-hand side measures how well the model fits the training data, the second term acts as a penalty to control model complexity and prevent overfitting, and the third term is a constant. Maximizing this function yields the optimal hyperparameter estimates.

[0053] The desired improvement acquisition function used in Bayesian optimization is: Where f+ represents the maximum glass transition rate observed in the current training set; (x) and (x) are the predicted mean and predicted standard deviation at the query point, respectively, given by the first two equations; Let represent the cumulative distribution function and probability density function of the standard normal distribution, respectively. The first term on the right-hand side of the equation encourages development behavior when the predicted mean is higher than the current optimal value, and the second term encourages exploration behavior when the predicted standard deviation is large. The input point corresponding to the global maximum value of this function is the optimal operating parameter recommended for the next batch.

[0054] The incremental update expression for the covariance matrix is ​​used to support efficient iteration of the model as data accumulates: Where the subscript t represents the current training set size, x t+1 For the newly added sample points, K t The old covariance matrix, This represents the covariance vector between the new sample point and each sample in the old training set. This incremental update avoids the computational burden of recalculating the entire covariance matrix after each new sample, enabling the model to maintain real-time responsiveness in large-scale continuous production scenarios. All formulas in this step are executed simultaneously, collectively realizing the transformation of the water quenching process from manual experience-based operation to data-driven intelligent decision-making.

[0055] S5: Same as in Example 1.

[0056] S6: Same as in Example 1.

[0057] S7: Same as in Example 1.

[0058] Comparative example: The comparative example uses a traditional CaO-SiO2-CaF2 system high-titanium steel protective slag, comprising, by weight: 27 parts CaO, 35 parts SiO2, 10 parts CaF2, 5 parts Al2O3, 8 parts Na2O, 5 parts MgO, 3 parts Fe2O3, and 7 parts C. The preparation method is a conventional process, with the specific steps as follows: S1: Weigh CaO, SiO2, CaF2, Al2O3, Na2O, MgO, Fe2O3 and C according to the above ratio and load them into a planetary ball mill. Mix at 250 r / min for 30 min.

[0059] S2: Transfer the well-mixed powder into a graphite crucible, place it in a resistance furnace and heat it to 1450℃ at 10℃ / min. After the material is completely melted, keep it at that temperature for 30 min.

[0060] S3: After the heat preservation is completed, remove the crucible from the furnace and pour the high-temperature melt directly onto a flat steel plate to cool and solidify naturally. Remove the slag after it has cooled to room temperature.

[0061] S4: First, crush the solidified slag blocks with a hammer to a particle size of less than 10 mm, then coarsely crush them with a jaw crusher, finely grind them with a disc pulverizer, and pass them through a 200-mesh standard sieve. The material passing through the sieve is directly sealed and packaged to obtain the finished product.

[0062] The comparative example used natural cooling and solidification after smelting instead of water quenching, so it did not involve the control and measurement of the glass transition rate of the slag, nor did it use any machine learning-assisted optimization methods.

[0063] A comparison of the slag film structure and heat transfer performance of different protective slags is shown in the figure. Figure 2 Table 2 compares the effects of different protective slags on continuous casting. Figure 3 ,Depend on Figure 2 and Figure 3It can be seen that both Examples 1 and 2 used a fluorine-free CaO-Al2O3-TiO2-based formulation, with slag film thicknesses reaching 2.73 cm and 2.98 cm, respectively, significantly higher than the 1.85 cm of the comparative example; surface roughnesses were 18.37 μm and 19.52 μm, respectively, much higher than the 11.24 μm of the comparative example; apparent porosity and true porosity were also significantly better than those of the traditional fluorine-containing slag system. In terms of heat transfer performance, the steady-state heat flux densities of Examples 1 and 2 were 0.558 MW / m and 0.536 MW / m, respectively, and the apparent comprehensive thermal conductivity was 1.09 W / m², respectively. -1 K -1 and 1.07 Wm -1 K -1 Both were significantly lower than the comparative figures of 0.742 MW / m and 1.25 W / m. -1 K -1 This indicates that both possess excellent heat transfer suppression capabilities.

[0064] In Example 2, based on Example 1, an online optimization module for water quenching based on Gaussian process regression was introduced. The glass transition rate of the slag increased from 84.7% to 94.3%, the slag film thickness increased from 2.73 cm to 2.98 cm, the surface roughness increased from 18.37 μm to 19.52 μm, the true porosity increased from 5.08 vol% to 5.29 vol%, the steady-state heat flux density further decreased to 0.536 MW / m, and the thermal conductivity decreased to 1.07 W / m². -1 K -1 The heat transfer control capability was further enhanced. Compared with the comparative example, the surface crack index of the billet in Example 2 decreased from 8.6 m / m to 2.8 m / m, a reduction of 67.4%; the heat flow fluctuation rate of the crystallizer narrowed from 9.6% to 2.9%; the number of steel leakage alarms decreased from 5 times / thousand tons to 0 times; and the slag consumption decreased from 0.67 kg / t to 0.48 kg / t.

[0065] The above results show that the fluorine-free protective slag provided by the present invention achieves a better heat transfer control effect while ensuring good lubrication, and has significant engineering value for improving the surface quality of high-titanium steel billets.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the term "include" or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the inclusion of a defined element by a statement does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0067] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A fluorine-free continuous casting protective slag for high-titanium steel with strong heat transfer control capability, characterized in that: include: CaO: 28 parts by weight, Al2O 3: 26 parts by weight, TiO 2: 8 parts by weight, Na₂O: 10 parts by weight, SiO 2: 5 parts by weight, BaO: 12 parts by weight, B2O 3: 11 parts by weight.

2. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: Includes the following steps: S1: Calculate and weigh the chemically pure raw materials CaO, Al2O3, TiO2, SiO2, B2O3, Na2CO3 and BaCO3 according to the formula ratio; before weighing, dry all raw materials in an oven at 105~110℃ for 2~4 hours to remove free moisture; S2: Load all the weighed materials into the planetary ball mill jar, control the ball-to-material ratio to be 1, and dry ball mill for 30 minutes to make each component reach a fully uniform powder state. S3: Transfer the mixed powder to a high-purity graphite crucible, place it in a medium-frequency induction furnace, and heat it to 1400℃ at a heating rate of 5~8℃ / min. After the material is completely melted, keep it at that temperature for 20 minutes. During the holding period, stir it gently with a corundum rod every 5 minutes to ensure that the melt is uniform. S4: After the heat preservation is completed, the crucible is immediately removed. The optimal pouring parameters are recommended based on the real-time data of the sensor using the Gaussian process regression model. The operator then continuously pours the high-temperature melt into a room-temperature deionized water bath for rapid cooling and quenching to obtain glassy slag. After water quenching, the glass transition rate and corresponding parameters are fed back to the model through rapid XRD detection to achieve iterative optimization for subsequent batches. S5: Remove the water-quenched slag blocks, spread them on a stainless steel tray, and dry them in a 105℃ forced-air drying oven for 4-6 hours to remove the attached water. S6: The dried slag is first coarsely crushed to a particle size of less than 5mm using a jaw crusher, and then transferred to a disc crusher or vibrating mill for fine grinding to further refine the particles; S7: Pass the finely ground powder through a 200-mesh standard sieve, take the sieve material and place it in a V-type mixer to continue mixing for 15 minutes to eliminate local segregation that may be caused by grinding, and then seal and package it to obtain the finished product.

3. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S1, all raw materials must be placed separately in a forced-air drying oven before weighing and continuously heated and dried for 2 to 4 hours in a temperature range of 105 ℃ to 110 ℃ to remove free moisture adsorbed on the surface and inside of the particles.

4. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S2, zirconia balls are added in a ratio of 1:2 to 1:3 between the total mass of powder and the mass of grinding media. After the container is closed, the ball mill is started and dry mixing and grinding are performed at a spindle speed of 250 r / min to 300 r / min. The total grinding time is 30 min, and the mill is paused every 10 min to prevent the container from overheating.

5. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S3, after closing the furnace door, the heating program is started, and the heating rate is controlled to be maintained between 5 ℃ / min and 8 ℃ / min. When the temperature reaches 1400 ℃, the heat preservation mode is switched and the temperature is maintained for 20 min. During the heat preservation period, every 5 min, a corundum rod with a diameter of about 6 mm is inserted vertically from the top of the furnace into the crucible molten pool and gently stirred once. Each stirring lasts for about 30 seconds.

6. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S4, the crucible is slowly tilted so that the high-temperature melt flows into the water in a continuous thin stream. The entire tilting process is completed within 5 to 8 seconds. The melt is rapidly cooled and quenched upon entering the water, resulting in transparent or translucent glassy slag particles.

7. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S4, the crucible is immediately removed and transferred to the water quenching station after the heat preservation is completed, and the online optimization module based on Gaussian process regression is started. This module constructs an initial training set with at least 20 sets of historical data, including melt temperature, pouring time, water temperature, liquid surface distance and ambient humidity as inputs and the glass transition rate of the slag measured by XRD as output. The hyperparameters of the Matern 5 / 2 kernel function are trained by the logarithmic marginal likelihood function. Then, the glass transition rate is predicted by the posterior mean formula and the posterior variance formula for the query point consisting of the real-time working condition transmitted by the current sensor and the operation parameters to be optimized. Then, the expected improvement formula is used as the acquisition function to globally optimize in the two-dimensional operation space. The optimal pouring duration and liquid flow impact angle are displayed on the operation terminal for the operator to perform water quenching. After water quenching, the measured glass transition rate and the current operation parameters constitute a new sample point. The model iteration is realized by the covariance matrix incremental update formula.

8. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S5, the water-quenched slag particles are scooped out of the water tank with a stainless steel strainer, drained of a large amount of free water, and then evenly spread on a stainless steel tray. The spreading thickness should not exceed 3 cm to ensure uniform drying in the future. The tray is placed in a forced-air drying oven and the temperature is set to 105 degrees Celsius for continuous drying for 4 to 6 hours.

9. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S6, the discharge gap of the crusher is adjusted so that the maximum size of the crushed particles is controlled below 5 mm. After coarse crushing, the obtained particles are transferred to a disc crusher for medium crushing and fine grinding. The disc spacing is gradually tightened so that the material undergoes splitting, grinding and shearing in sequence, and the final discharge particle size reaches the level of all passing through a 150-mesh standard sieve.

10. The method for preparing a high-titanium steel fluorine-free continuous casting protective slag with strong heat transfer control capability according to claim 1, characterized in that: In step S7, the mixer speed is set to 12 to 15 revolutions per minute, and the mixture is continuously mixed for 15 minutes to eliminate local segregation caused by density differences during the previous crushing and grinding process. After mixing, the machine is stopped and the discharge port is opened. The powder is then passed through a 200-mesh standard vibrating screen for final sieving. The material under the screen is taken as the qualified finished product. The material over the screen is returned to the disc pulverizer for fine grinding and then sieved again. The qualified powder is placed in a double-layer moisture-proof plastic bag, the air inside the bag is removed, and the bag is heat-sealed. It is then placed in a sealed metal barrel for storage in the dark and dry place. A label is attached to the outside of the barrel indicating the sample number, production date, and sieving particle size information. This completes all the preparation steps of the high-titanium steel fluorine-free continuous casting protective slag.

Images