Intelligent bottom blowing stirring control method and device for LF refining furnace and storage medium

Abstract

The invention discloses an LF refining furnace intelligent bottom blowing stirring control method and device, a storage medium method and device and a storage medium. Obtaining sensing data about the target LF refining furnace, wherein the sensing data comprises LF furnace molten steel liquid level image data and LF furnace bottom blowing gas detection data; performing feature extraction processing on the image data and the detection data based on a pre-trained neural network model to obtain fusion features between the image data and the detection data; according to the fusion features, an LF furnace bottom blowing control strategy is obtained, and the control strategy is used for describing setting of bottom blowing data; and according to the bottom blowing control strategy, a fuzzy logic controller is established, the flow and pressure of bottom blowing gas are adjusted, and accurate control over bottom blowing stirring of the LF refining furnace is achieved. By means of the method, the device and the storage medium, accurate control over bottom blowing gas stirring in the ladle refining process can be achieved, the molten steel quality is improved, the production cost is reduced, and the important practical application value is achieved.

Description

Technical Field

[0001] This application belongs to the field of iron and steel metallurgical control technology, and in particular relates to an intelligent bottom blowing stirring control method, device and storage medium for LF refining furnace. Background Technology

[0002] The primary function of the LF furnace is to alloy molten steel, adjust its composition, perform deep desulfurization and inclusion removal, and regulate temperature through heating to coordinate the rhythm of the primary refining furnace and the continuous casting machine. Currently, over 90% of steel grades require LF refining. Some high-end steel grades generally require deep deoxidation, deep desulfurization, and inclusion removal in the LF, resulting in longer refining times. The refining effect directly impacts steel quality and production costs. Bottom-blowing agitation technology is a crucial technique in the secondary refining process of the LF furnace, significantly improving steel quality. Currently, most bottom-blowing agitation control systems are manually operated, with workers visually observing the molten steel level in the ladle and controlling the flow rate based on experience. This increases the error rate and reduces efficiency. This outdated monitoring and management mechanism is far from meeting the demands of production and technological development. Therefore, there is an urgent need to develop a method for accurately and efficiently automatically adjusting bottom-blowing agitation in the ladle to improve the refining level of the LF furnace and effectively enhance the intelligence and automation of bottom-blowing agitation control in the LF furnace. Summary of the Invention

[0003] This invention provides at least one intelligent bottom-blowing stirring control method, device, and storage medium for LF refining furnaces.

[0004] The first aspect of this application provides an intelligent bottom-blowing stirring control method for an LF refining furnace, comprising: acquiring sensing data about a target LF refining furnace, the sensing data including LF furnace molten steel level image data and LF furnace bottom-blowing gas detection data; performing feature extraction processing on the image data and detection data based on a pre-trained neural network model to obtain fusion features between the image data and detection data; obtaining an LF furnace bottom-blowing control strategy based on the fusion features, the control strategy being used to describe the setting of the bottom-blowing data; and establishing a fuzzy logic controller based on the bottom-blowing control strategy to adjust the flow rate and pressure of the bottom-blowing gas, thereby achieving precise control of the bottom-blowing stirring of the LF refining furnace.

[0005] In one embodiment, the acquisition of sensing data regarding the target LF refining furnace includes LF furnace molten steel level image data and LF furnace bottom-blown gas detection data. This includes: acquiring the LF furnace molten steel level image using an industrial high-temperature resistant camera and a dedicated furnace infrared lens; acquiring the LF furnace molten steel level image using a water-cooled probe installed in the gap between the furnace cover water-cooling pipes; and acquiring the LF furnace molten steel level image using industrial Ethernet protocol transmission. The LF furnace bottom-blown gas detection data includes the type, flow rate, and pressure of the bottom-blown gas, as well as the multi-pipeline configuration.

[0006] In one embodiment, the pre-trained neural network model includes a visual encoding module and a text encoding module. The step of extracting features from the image data and the detection data based on the pre-trained neural network model to obtain image features corresponding to the image data and text features corresponding to the detection data includes: inputting the image data into the visual encoding module of the neural network model to obtain the image features; and inputting the detection data into the text encoding module of the neural network to obtain the text features.

[0007] In one embodiment, the step of obtaining the bottom blowing control strategy of the LF furnace based on the fusion characteristics, wherein the control strategy is used to describe the setting of the bottom blowing data, includes: when the fusion characteristics are in the target refining state stage and the current refining target of the target LF furnace has not been completed, determining that the bottom blowing control setting of the target LF furnace remains unchanged; when the fusion characteristics are in the target refining state stage and the current refining target of the target LF furnace has been completed, determining that the bottom blowing control setting of the target LF furnace is the value of the next stage; wherein the stage is: breaking the top, uniform temperature sampling, adding refining slag, heating and raising the temperature, alloy fine-tuning and soft blowing.

[0008] In one embodiment, after the step of obtaining an LF furnace bottom blowing control strategy based on the fusion features, wherein the control strategy is used to describe the setting of bottom blowing data, the method further includes: based on the LF furnace bottom blowing control strategy, the control strategy includes refining targets for the LF furnace; based on the LF furnace bottom blowing control strategy, the control strategy takes into account the production and operation status of the LF furnace permeable bricks.

[0009] In one embodiment, the step of establishing a fuzzy logic controller based on the bottom-blowing control strategy to adjust the flow rate and pressure of the bottom-blowing gas and achieve precise control of the LF furnace bottom blowing includes: determining the specific range of the flow rate based on the flow rate data of the bottom-blowing gas; determining the specific range of the pressure based on the pressure data of the bottom-blowing gas; constructing a fuzzy control rule base, including: constructing a fuzzy control rule base based on the specific range of the flow rate and the specific range of the pressure, combined with the LF refining process requirements; performing fuzzification processing on the acquired LF furnace bottom-blowing gas pressure and flow rate data, including: using one of Gaussian membership functions, trapezoidal membership functions, or triangular membership functions to convert the LF furnace bottom-blowing gas pressure and flow rate data into corresponding fuzzy sets; obtaining the fuzzy quantities of the LF furnace bottom-blowing pressure and flow rate values ​​that need to be adjusted through fuzzy inference, including: performing fuzzy inference using the Sugenno inference method based on the fuzzy control logic rule base and the fuzzified bottom-blowing gas flow rate fuzzy set and pressure fuzzy set, to obtain the LF furnace bottom-blowing pressure and flow rate values ​​that need to be adjusted. Fuzzy quantities of bottom-blown gas pressure and flow rate; defuzzification processing of the adjustment fuzzy quantities obtained from fuzzy inference, including: using either maximum membership degree or weighted average division to convert the adjustment fuzzy values ​​obtained from fuzzy inference into precise flow control values ​​and pressure control values.

[0010] In one embodiment, the step of acquiring sensing data about the target LF furnace includes: controlling a data acquisition device to perform a preset data acquisition task, the data acquisition device being equipped with multiple sensors, including an image sensor, a pressure sensor, and a flow sensor; acquiring image data of the LF furnace refining liquid level based on the image sensor, and acquiring pressure data based on the pressure sensor and flow data based on the flow sensor.

[0011] A second aspect of this application provides an intelligent bottom-blowing stirring control device for an LF furnace, comprising: a data acquisition module for acquiring sensing data of the LF refining state to be controlled, the sensing data including LF furnace molten steel level image data and LF furnace bottom-blowing gas detection data; a data feature processing module for performing feature extraction processing on the image data and detection data based on a pre-trained neural network model to obtain fusion features between the image data and detection data; a control decision module for obtaining an LF furnace bottom-blowing control strategy based on the fusion features, the control strategy being used to describe the setting of the bottom-blowing data; and a control execution module for establishing a fuzzy logic controller based on the bottom-blowing control strategy to adjust the flow rate and pressure of the bottom-blowing gas, thereby achieving precise control of the LF furnace bottom-blowing stirring.

[0012] The third aspect of this application provides a storage medium for intelligent bottom blowing and stirring control of an LF furnace, which stores programs and instructions, and the above-mentioned intelligent bottom blowing and stirring control method for an LF furnace is implemented when the programs and instructions are executed.

[0013] In some embodiments of this application, the technical solutions involve acquiring image data of the molten steel level and detection data of the bottom blowing gas of the target LF furnace; performing feature extraction processing on the image data and detection data based on a pre-trained neural network model to obtain fusion features between the image data and detection data; obtaining the bottom blowing control strategy of the LF furnace based on the fusion features; and establishing a fuzzy logic controller based on the bottom blowing control strategy to adjust the flow rate and pressure of the bottom blowing gas, thereby achieving precise control of the bottom blowing of the LF furnace. Compared with traditional methods, this method can effectively control the bottom blowing stirring of the LF furnace and improve the refining level of the LF furnace.

[0014] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this application. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 is a schematic flowchart of an embodiment of the intelligent bottom blowing stirring control method for an LF furnace according to this application; Figure 2 is a schematic diagram of the deep learning process of an LF furnace intelligent bottom blowing stirring control method according to an embodiment of this application; Figure 3 is a fuzzy control structure diagram of an embodiment of the intelligent bottom blowing stirring control method for an LF furnace according to this application; Figure 4 is a block diagram of an LF furnace intelligent bottom blowing stirring control device according to an embodiment of this application; Figure 5 is a schematic diagram of the storage medium structure for intelligent bottom blowing stirring control of an LF furnace according to an embodiment of this application.

[0017] The technical solutions in the embodiments of this application will be described in full and clearly below with reference to the accompanying drawings and specific examples.

[0018] It should be noted that, in non-conflicting scenarios, the following embodiments and features can be combined with each other or implemented individually; and, based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the following description, specific details such as interfaces, structures, technologies, and systems are presented for illustrative purposes rather than limiting purposes in order to thoroughly explain this application.

[0020] In the following description, the term "intelligent" is used to indicate that this method differs from traditional methods; "gas" refers to either nitrogen or argon.

[0021] Figure 1 is a schematic flowchart of an exemplary embodiment of the intelligent bottom blowing stirring control method for LF furnace of this application, including the following steps: Step S110: Obtain sensing data about the target LF furnace, including LF furnace molten steel level image data and LF furnace bottom blowing gas detection data.

[0022] Sensing data refers to the data that can be collected for the bottom-blown gas of the LF furnace to be controlled and the image data of the molten steel level in the LF furnace in the application scenario of bottom-blown gas control in the refining process of the LF furnace.

[0023] Specific acquisition methods may include, but are not limited to, using acquisition equipment. For example, acquiring images of the molten steel level in an LF furnace may involve using a high-temperature resistant industrial camera and a dedicated furnace infrared lens.

[0024] The acquisition of the LF furnace molten steel level image data includes the following steps: The S1101 camera selection utilizes an industrial high-temperature resistant camera and a dedicated furnace infrared lens. The industrial high-temperature resistant camera is used to acquire, store, and process thermal images and temperature arrays of the furnace or molten steel, while the infrared lens enables stable operation even in harsh high-temperature environments.

[0025] S1102, a fixed endoscopic high-temperature industrial camera, is primarily used for image inspection in high-temperature environments. It is installed in the gap between the cooling water pipes of the furnace cover and uses water cooling. The probe is made of double-layered stainless steel, with water directly cooling the interlayer. Compressed air is used to clean the camera and lens, preventing dust from adhering to the lens. The camera lens can be directly inserted into the furnace to observe the LF furnace condition.

[0026] S1103, data acquisition, uses an industrial high-temperature resistant camera and a special furnace infrared lens to acquire images of the furnace interior, and transmits them via network cable to a color monitor for image display and video recognition.

[0027] S1104, In this embodiment, the protective measures for the industrial high-temperature camera and the infrared lens for the special furnace include: 1) Water circuit: Cooling water flows through the entire embedded part base plate and is circulated out from the outlet; 2) Gas path: Prevents backflow of flames inside the furnace and eliminates the corrosion of the lens by high-temperature dust inside the furnace; 3) Circuit: AC220V±10% power supply is used. The external power supply is directly connected to the electrical control box and outputs after being controlled by a switch. 4) Underpressure alarm: An alarm will be triggered when the on-site air pressure is lower than the set value; 5) Overtemperature alarm: An alarm will be triggered when the on-site temperature is higher than the set temperature.

[0028] The LF furnace bottom-blowing gas detection data includes the type of bottom-blowing gas, the piping layout, pressure, and flow rate. Common bottom-blowing gases include argon and nitrogen. These gases can effectively agitate the molten steel, promoting the flotation of impurities and thus improving the purity of the steel. The piping layout for the bottom-blowing gas needs to take into account factors such as the furnace structure and gas flow characteristics to ensure that the gas is blown into the molten steel evenly and effectively.

[0029] Furthermore, generally, nitrogen and argon are used simultaneously for bottom blowing, with two pipelines supplying each gas. The pressure and flow rate of the bottom blowing gas are key parameters affecting its stirring effect. The pressure and flow rate of the bottom blowing gas should be set appropriately based on actual production needs and equipment conditions.

[0030] Other test data for LF furnace bottom blowing include parameters of the permeable bricks, specifically: number of permeable bricks, diameter of the top surface of the permeable brick (mm), distance from the center of the permeable brick to the center of the bottom of the furnace (mm), and the angle between the line connecting the centers of the permeable bricks to the major axis of the bottom of the furnace. The parameters of the permeable bricks affect the flow rate and pressure of the bottom-blown gas; therefore, the influence of these parameters must be considered when testing the flow rate and pressure of the bottom-blown gas in the LF furnace.

[0031] Step S120: Based on the pre-trained neural network model, feature extraction processing is performed on the image data and detection data to obtain the fusion features between the image data and detection data.

[0032] Furthermore, the pre-trained neural network model is used to process data obtained through various techniques such as image recognition, data detection, and data analysis to achieve precise control over the refining stirring effect of the LF furnace. For example, the neural network model is trained using a convolutional neural network (CNN), which can process various types of information, including images and numbers.

[0033] Specifically, the following example illustrates the use of image data of molten steel level in the LF furnace and bottom blowing data from the LF furnace as the sensing data. After inputting these data into a convolutional neural network (CNN) model, the model performs feature extraction processing on both the image features and the detection data, obtaining image features corresponding to the image data and digital features corresponding to the detection data. Then, the image features and digital features are fused to obtain fused features. The neural network model then performs model calculations and inferences based on these fused features, and uses the calculation and inference results to determine the refining stage of the LF furnace, obtaining the decision result corresponding to the fused features, thereby determining the refining stage of the LF furnace.

[0034] Furthermore, in the LF furnace refining bottom blowing control method of this application, the neural network model can be trained and reasoned based on fused features to determine the refining state stage of the LF furnace; while the traditional LF furnace refining state is usually determined manually, based on refining time characteristics and visual observation of the molten steel surface state. This method requires human experience and is less efficient than the LF furnace bottom blowing gas control method of this application.

[0035] Based on the above embodiments, the above convolutional neural network (CNN) automatically learns the feature representation of the input data through multi-layer convolution, activation, pooling and fully connected operations, thereby achieving efficient image recognition, classification and other tasks. The convolutional neural network (CNN) adopts, for example, ResNet or DenseNet, which can effectively extract complex features due to their deep structure and residual learning.

[0036] Furthermore, the specific steps are as follows: S1201, Data Collection and Preprocessing: This involves collecting and preparing training data, including LF furnace molten steel surface images and corresponding labels, as well as LF bottom-blowing gas detection data. The labels represent different stages in the LF furnace refining process, such as: top breaking, isothermal sampling, adding refining slag, heating and temperature increase, alloy fine-tuning, and soft blowing. The liquid level height and fluctuation characteristics differ significantly between these stages. Image rotation, scaling, and cropping are used to increase the data diversity of the LF furnace molten steel surface images and reduce the risk of overfitting.

[0037] S1202, Feature Extraction: This section uses convolutional layers to extract image features of the molten steel surface in the LF furnace. In a convolutional layer, a set of learnable filters performs convolution operations on the input data. Each filter generates a new feature map, capturing different features in the input data, such as edges, textures, and shapes. In a Convolutional Neural Network (CNN), filters slide across the input data (such as an image), performing convolution operations at each location to generate a new feature map. This feature map can be seen as the features extracted by the filters from the input data. For example, one filter might be sensitive to image edges, while another might be sensitive to color changes. The size and number of filters are adjustable hyperparameters. Generally, more filters allow the network to learn more features, but also increase computational complexity and the risk of overfitting. The filter size determines the receptive field, i.e., the range of input data that each output unit can "see."

[0038] S1203 is the activation function, which performs non-linear processing on the data output of the convolutional layer, increasing the network's non-linear processing capabilities. The activation function used is the Softmax function: the Softmax function is commonly used in multi-class classification tasks to transform the output into a probability distribution. The formula for the Softmax function is f(x) = exp(x...). i ) / Σexp(x j ), where i and j represent different categories.

[0039] S1204, pooling layer operation, reduces the spatial dimensionality of the feature map while preserving important feature information. This helps reduce computational cost and the risk of overfitting. Max pooling is used, dividing the input feature map into multiple non-overlapping rectangular regions, and then extracting the maximum value from each region as the output. This reduces the spatial size of the feature map while retaining the maximum response within local regions.

[0040] S1205, the fully connected layer, connects the output of the pooling layer to a fully connected layer for classification. The output of the pooling layer is flattened and then used as the input to the fully connected layer. The output of the fully connected layer is then used for classification to obtain the predicted LF furnace stage identification result.

[0041] S1206, the loss function, uses Softmax Cross-Entropy Loss to calculate the difference between the predicted probability distribution and the true distribution. The formula is: Softmax Cross-Entropy = -Σy_true *log(softmax(y_pred)), where y_true is the index of the true LF furnace stage, and y_pred is the prediction result.

[0042] S1207 is an optimization algorithm that uses the Adam (Adaptive Moment Estimation) algorithm. Adam adaptively adjusts the learning rate by calculating the first moment (mean) and second moment (uncentered variance) of the gradient, making it a very effective optimization algorithm. Its update rule is as follows: Calculate the gradient value of each parameter; Update the first moment estimate: Multiply the current gradient value by a decay factor and then add a weighted average of the historical gradient values; Update the second moment estimate: Multiply the square of the current gradient value by a decay factor and then add a weighted average of the squared historical gradient values; Calculate the updated parameter value: Subtract the first moment estimate from the original parameter value, divide by the square root of the second moment estimate, and then multiply by the learning rate. In this way, the adaptive moment estimation algorithm can dynamically adjust the learning rate based on the historical gradient information of each parameter, thereby improving the convergence speed and reducing the risk of overfitting.

[0043] S1208, Model Evaluation. Model evaluation refers to the process of quantitatively analyzing the performance of a Convolutional Neural Network (CNN) model. Its purpose is to determine the model's accuracy, reliability, and efficiency on a specific task. First, a confusion matrix is ​​calculated to show the comparison between the model's predictions and actual results at each LF furnace refining stage. Precision, recall, and F1 score are calculated using the confusion matrix. Second, an ROC curve is plotted to describe the relationship between the true positive rate (TPR) and false positive rate (FPR) at different thresholds, measuring the model's overall performance. Then, cross-validation is performed, dividing the dataset into multiple subsets. One subset is used as the test set, and the remaining subsets are used as the training set, repeating the training and testing process. Finally, the model's average performance is calculated to evaluate its generalization ability. Next, by observing the performance difference between the training and test sets, it can be determined whether the model has overfitting or underfitting issues. Finally, cross-entropy loss is used to calculate the difference between the model's predicted and actual values. By observing the value of the loss function, the degree of optimization during the training process can be understood.

[0044] Step S130: Based on the fusion characteristics, obtain the LF furnace bottom blowing control strategy, which is used to describe the setting of bottom blowing data.

[0045] Based on the above embodiments, the application scenario of LF furnace bottom blowing stirring control will be described as an example. It should be noted that the LF furnace bottom blowing stirring control system of this application can continuously acquire and control the sensing data of LF furnace bottom blowing stirring, and therefore will generate different judgment results based on different LF furnace refining stages. The LF furnace bottom blowing control system can save these judgment results. According to the control rules corresponding to different stages, the corresponding control strategy is determined.

[0046] For example, the bottom blowing control rules for the refining stage of the LF furnace can be shown in Table 1 below: Refining stage Flow rate (L / min) Pressure (MPa) Time (s) Breakthrough stage 900 0.8 6 Temperature equalization sampling stage 150 0.15 210 Refining slag stage 250 0.2 233 Heating and temperature rise stage 600 0.4 1100 Alloy Fine-tuning Step 200 0.3 180 soft blow stage 150 0.15 220

[0047] For example, as shown in the table, this application acquires various sensory data on the bottom blowing agitation of the LF furnace to be controlled, and inputs the sensory data into a pre-trained convolutional neural network (CNN) model for feature extraction, obtaining fused features composed of features from image sensory data and numerical sensory data, providing comprehensive identification results for the refining stage of the LF furnace; the model analyzes and understands the descriptive features of the refining stage of the LF furnace, thereby finding the target stage result that matches its textual features in the identification results and providing feedback, facilitating the use of this method to control the bottom blowing of the LF furnace, thereby achieving accurate and efficient bottom blowing agitation control of the LF refining furnace.

[0048] Based on the above embodiments, this application embodiment obtains an LF furnace bottom blowing control strategy according to the fusion characteristics. Specifically, the method of this embodiment includes the following steps: S1301, based on the fused features, a decision tree model is used to generate the LF furnace bottom blowing control strategy. The decision tree, through recursive segmentation of the feature space, can automatically select the features with the greatest predictive power for the output variables and make decisions accordingly. First, the correlation between the features is analyzed to determine which features can be merged; second, PCA or similar techniques are applied to fuse features, reducing dimensionality and forming a new feature set; then, a suitable decision tree algorithm is selected, such as CART, ID3, C4.5, etc.; the decision tree model is trained using a training dataset, with the fused features as input; finally, decision rules are extracted from the trained decision tree, and the control parameter adjustment strategy for LF furnace bottom blowing is defined according to the decision rules.

[0049] S1302 converts the control strategy into instructions that the control system can execute, monitors the LF furnace bottom blowing process in real time, and adjusts the control strategy according to the actual effect; S1303 updates training data periodically to adapt to process changes, and retrains the model when data and processes are updated.

[0050] Step S140: Based on the bottom blowing control strategy, a fuzzy logic controller is established to adjust the flow rate and pressure of the bottom blowing gas, thereby achieving precise control of the bottom blowing of the LF furnace.

[0051] Furthermore, this embodiment includes the following steps: S1401 Real-time monitoring of bottom-blown gas type, flow rate, and pressure: During the refining process, the flow rate of bottom-blown gas is monitored in real time by flow sensors distributed at the bottom of the LF furnace. The collected flow data is converted into a flow data stream Q(t), where t is time and Q(t) represents the flow rate value collected at time t. S1402, using a fuzzy logic control algorithm to dynamically adjust flow rate and time: Based on real-time monitored flow data, a fuzzy logic control algorithm is used to dynamically adjust the gas flow rate and time to ensure the efficiency and stability of the bottom blowing stirring process. Further, the fuzzy logic control algorithm specifically includes: Fuzzy logic control input: The flow data stream Q(t) is used as the input to the fuzzy logic control algorithm. The flow error E(t) and the flow change rate ΔE(t) are set: E(t) = Qset - Q(t); ΔE(t) = E(t) - E(t-1); where Qset is the set target flow. Fuzzy logic inference: A fuzzy control rule base is set, and fuzzy sets (such as "high", "medium", "low") of flow error E(t) and flow change rate ΔE(t) are defined. Based on the real-time flow data Q(t), flow error E(t), and flow change rate ΔE(t), the flow adjustment amount is calculated through the fuzzy inference system.

[0052] Furthermore, a fuzzy logic control algorithm is used to dynamically adjust the pressure and time: based on real-time monitored pressure data, a fuzzy logic control algorithm is used to dynamically adjust the gas pressure and time to ensure the efficiency and stability of the bottom-blowing stirring process. Furthermore, the fuzzy logic control algorithm specifically includes: Fuzzy logic control input: The pressure data stream P(t) is used as the input to the fuzzy logic control algorithm. The pressure error Ep(t) and the pressure change rate ΔEp(t) are set: Ep(t) = Pset - P(t); ΔEp(t) = Ep(t) - Ep(t-1), where Pset is the set target flow rate; Fuzzy logic reasoning: Set up a fuzzy control rule base and define fuzzy sets (such as "high", "medium", "low") of pressure error Ep(t) and pressure change rate ΔEp(t); Based on real-time pressure data Qp(t), pressure error Ep(t), and pressure change rate ΔEp(t), the flow adjustment amount is calculated using a fuzzy inference system. Fuzzy control output: Defuzzify the output of the fuzzy inference system to obtain the specific flow adjustment amount △P(t); Dynamic adjustment of bottom blowing time: Based on the result of flow adjustment and combined with production needs, dynamically adjust the bottom blowing time, set a fuzzy control rule base, and define the fuzzy set of flow change and time adjustment; calculate the time adjustment amount △T(t) through the fuzzy inference system and dynamically adjust the bottom blowing time.

[0053] As shown in Figure 2, according to one aspect of this application, Figure 2 is a schematic diagram of the deep learning process of an LF furnace intelligent bottom blowing stirring control method according to an embodiment of this application. The deep learning adopts a convolutional neural network (CNN). The CNN automatically learns the feature representation of the input data through multi-layer convolution, activation, pooling and fully connected operations, thereby achieving efficient image recognition, classification and other tasks. The CNN adopts, for example, ResNet or DenseNet. These networks can effectively extract complex features due to their deep structure and residual learning. The deep learning steps in Figure 2 are further as follows: Step S210: Data acquisition, real-time acquisition of LF furnace liquid level image data and bottom blowing detection data; Step S220: Data preprocessing, including grayscale processing, noise reduction, and edge detection of the acquired image data; Step S230, Feature extraction, using convolutional layers to extract features; Step S240, Feature analysis, using multi-layer convolution and pooling for feature analysis; Step S250: Optimize the algorithm by using a Softmax classifier and a fully connected layer. Step S260, Model Evaluation: The model performance is evaluated using the confusion matrix, ROC, and cross-entropy loss calculation. Step S270: Data output, the recognition result is uploaded to the server.

[0054] Specifically, in step S210, training data is collected and prepared, including LF furnace molten steel surface image data and corresponding labels, as well as LF bottom-blowing gas detection data. The labels represent the textual representations of different stages in the LF furnace refining process, namely: top breaking, isothermal sampling, addition of refining slag, heating and temperature increase, alloy fine-tuning, and soft blowing, etc. The liquid level height and fluctuation type characteristics differ significantly between these different stages. Image rotation, scaling, and cropping are used to increase the data diversity of the LF furnace molten steel surface images and reduce the risk of overfitting.

[0055] Step S220: Data preprocessing, including grayscale processing, noise reduction, and edge detection of the acquired image data.

[0056] Step S230, Feature Extraction: Image features of the molten steel surface in the LF furnace are extracted using convolutional layers. In the convolutional layers, a set of learnable filters are used to perform convolution operations on the input data. Each filter generates a new feature map, which can capture different features in the input data, such as edges, textures, and shapes. In a Convolutional Neural Network (CNN), filters slide across the input data (such as an image), performing convolution operations at each location to generate a new feature map. This feature map can be seen as the features extracted by the filters from the input data. For example, one filter might be sensitive to image edges, while another filter might be sensitive to color changes. The size and number of filters are adjustable hyperparameters. Generally, the more filters there are, the richer the features the network can learn, but this also increases computational complexity and the risk of overfitting. The size of the filters determines the size of the receptive field, i.e., the range of input data that each output unit can "see."

[0057] Step S240, Feature Analysis: Pooling operations reduce the spatial dimensionality of the feature map while preserving important feature information. This helps reduce computational cost and the risk of overfitting. Max pooling is used to divide the input feature map into multiple non-overlapping rectangular regions, and then the maximum value is extracted from each region as the output. This reduces the spatial size of the feature map while preserving the maximum response within local regions. Softmax cross-entropy loss is used to calculate the difference between the predicted probability distribution and the true distribution. The output of the pooling layer is flattened and then used as the input to the fully connected layer. The output of the fully connected layer is then classified to obtain the predicted LF furnace stage recognition result.

[0058] Step S250: Optimize the algorithm using the Adam (Adaptive Moment Estimation) algorithm. The Adam update rule is as follows: Calculate the gradient value of each parameter; Update the first-order moment estimate: Multiply the current gradient value by a decay factor, and then add the weighted average of the historical gradient values; Update the second-order moment estimate: Multiply the square of the current gradient value by a decay factor, and then add the weighted average of the squared historical gradient values; Calculate the updated parameter values: Subtract the first-order moment estimate from the square root of the second-order moment estimate, and then multiply by the learning rate. In this way, the adaptive moment estimation algorithm can dynamically adjust the learning rate based on the historical gradient information of each parameter, thereby improving the convergence speed and reducing the risk of overfitting.

[0059] Step S260, Model Evaluation: First, a confusion matrix is ​​calculated to show the comparison between the model's predictions and actual results at each LF furnace refining stage. Precision, recall, and F1 score are calculated using the confusion matrix. Second, an ROC curve is plotted to describe the relationship between the true positive rate (TPR) and false positive rate (FPR) at different thresholds, measuring the model's overall performance. Then, cross-validation is performed, dividing the dataset into multiple subsets. One subset is used as the test set, and the remaining subsets are used as the training set, repeating the training and testing process. Finally, the model's average performance is calculated to evaluate its generalization ability. Next, by observing the performance difference between the model on the training and test sets, it can be determined whether the model has overfitting or underfitting issues. Finally, cross-entropy loss is used to calculate the difference between the model's predicted values ​​and actual values. By observing the value of the loss function, the degree of optimization during the training process can be understood.

[0060] Step S270: The recognition result is uploaded to the server. The transmission protocol is TCP / IP, and the network communication medium is optical fiber. The combination of TCP / IP protocol and optical fiber makes data transmission more efficient and reliable.

[0061] Figure 3 shows the steps to obtain the LF furnace bottom blowing control strategy based on the fusion features. The control strategy is used to describe the setting of bottom blowing data and the steps to establish a fuzzy logic controller based on the bottom blowing control strategy, adjust the flow rate and pressure of the bottom blowing gas, and realize the combination of precise control of LF furnace bottom blowing.

[0062] According to one aspect of this application, the step further comprises: Step S310: Based on the fused features, a decision tree model is used to generate the LF bottom blowing control strategy. First, the correlation between the features is analyzed to determine which features can be merged. Second, PCA or similar techniques are applied to fuse the features, reduce the dimensionality, and form a new feature set. Then, a suitable decision tree algorithm is selected, such as CART, ID3, C4.5, etc. The decision tree model is trained using a training dataset, with the fused features as input. Finally, decision rules are extracted from the trained decision tree, and the control parameter adjustment strategy for LF bottom blowing is defined according to the decision rules.

[0063] Step S320: The control strategy is converted into instructions that the control system can execute, the bottom blowing process of the LF furnace is monitored in real time, and the control strategy is adjusted according to the actual effect.

[0064] Step S330: Regularly update the training data to adapt to process changes. When the data and process are updated, retrain the model.

[0065] Step S340: Based on the bottom blowing control strategy, a fuzzy logic controller is established to adjust the flow rate and pressure of the bottom blowing gas, thereby achieving precise control of the bottom blowing of the LF furnace.

[0066] Step S350: Calculate the flow adjustment amount using a fuzzy inference system based on real-time flow data Q(t), flow error E(t), and flow change rate ΔE(t); calculate the flow adjustment amount using a fuzzy inference system based on real-time pressure data Qp(t), pressure error Ep(t), and pressure change rate ΔEp(t); calculate the time adjustment amount ΔT(t) using a fuzzy inference system, and dynamically adjust the bottom blowing time.

[0067] As shown in Figure 4, according to one aspect of this application, Figure 4 is a block diagram of an LF furnace intelligent bottom blowing stirring control device according to an embodiment of this application. The exemplary LF furnace intelligent bottom blowing stirring control device includes: a data acquisition module 410, a data feature processing module 420, a control decision module 430, and a control execution module 440.

[0068] Specifically, the data acquisition module 410 is used to acquire sensing data about the target LF furnace, including LF furnace molten steel level image data and LF furnace bottom blowing gas detection data.

[0069] The data feature processing module 420 is used to perform feature extraction processing on image data and detection data based on a pre-trained neural network model to obtain fused features between image data and detection data.

[0070] The control decision module 430 is used to obtain the LF furnace bottom blowing control strategy based on the fusion characteristics, and the control strategy is used to describe the setting of the bottom blowing data.

[0071] The control execution module 440 is used to establish a fuzzy logic controller according to the bottom blowing control strategy, adjust the flow rate and pressure of the bottom blowing gas, and realize precise control of the bottom blowing of the LF furnace.

[0072] This exemplary intelligent bottom-blowing stirring control device for an LF furnace senses data from the target LF furnace, including images of the molten steel surface and detection data of the bottom-blowing gas. This sensed data is then input into a pre-trained neural network model for feature extraction, yielding fused features between the image and detection data. This provides a comprehensive set of LF furnace bottom-blowing stirring control results. Based on these fused features, a control strategy describing the bottom-blowing data settings is derived. Finally, a fuzzy logic controller is established to adjust the flow rate and pressure of the bottom-blowing gas, achieving precise control of the bottom-blowing stirring in the LF refining furnace.

[0073] It should be noted that the apparatus designed in the above embodiments is for implementing the methods provided in the above embodiments. The specific functions of each module and the specific operation methods of each unit have been described in detail in the method embodiments, and will not be repeated here. In practical applications, the apparatus provided in the above embodiments can be allocated to different functional modules for execution as needed to complete all or part of the functions described above, and there is no limitation thereto.

[0074] As shown in Figure 5, according to one aspect of this application, Figure 5 is a schematic diagram of the storage medium structure for intelligent bottom blowing and stirring control of an LF furnace according to an embodiment of this application. The storage medium includes a memory 501 and a processor 502. The memory 501 stores programs, code, and instructions, and the processor 502 is used to execute the program instructions of 501 to realize the above-mentioned intelligent bottom blowing and stirring control of the LF furnace.

[0075] In one specific embodiment, the storage medium may be, but is not limited to, devices such as servers, microcomputers, embedded devices, tablet computers, and mobile laptops, and is not limited thereto.

[0076] In this exemplary LF furnace intelligent bottom blowing stirring control storage medium, data from the target LF furnace is sensed, including LF furnace molten steel level image data and LF furnace bottom blowing gas detection data. The sensed data is input into a pre-trained neural network model for feature extraction, obtaining fusion features between the image data and the detection data, providing a comprehensive set of LF furnace bottom blowing stirring control results. Based on the fusion features, a control strategy describing the bottom blowing data settings is obtained. Then, a fuzzy logic controller is established to adjust the flow rate and pressure of the bottom blowing gas, achieving precise control of the LF refining furnace bottom blowing stirring.

Claims

1. A method, apparatus, and storage medium for intelligent bottom-blowing stirring control of an LF furnace, characterized in that, The methods shown include: Acquire sensing data about the target LF furnace, including LF furnace molten steel level image data and LF furnace bottom blowing gas detection data; Based on a pre-trained neural network model, feature extraction processing is performed on image data and detection data to obtain fused features between image data and detection data; Based on the fusion characteristics, an LF furnace bottom blowing control strategy is obtained, which is used to describe the setting of bottom blowing data; Based on the bottom blowing control strategy, a fuzzy logic controller is established to adjust the flow rate and pressure of the bottom blowing gas, thereby achieving precise control of the bottom blowing of the LF furnace.

2. The method according to claim 1, characterized in that, The acquisition of sensing data about the target LF furnace includes LF furnace molten steel level image data and LF furnace bottom-blown gas detection data. Specifically, the acquisition of the LF furnace molten steel level image data involves using an industrial high-temperature resistant camera and a dedicated furnace infrared lens; the LF furnace molten steel level image is obtained by installing the camera in the gap of the furnace cover water-cooling pipes, using a water-cooling method. The probe is made of double-layered stainless steel; the LF furnace molten steel level image is transmitted using the industrial Ethernet protocol; and the LF furnace bottom-blown gas detection data includes the type, flow rate, and pressure of the bottom-blown gas, as well as the multi-pipeline configuration.

3. The method according to claim 1, characterized in that, The step of performing feature extraction processing on image data and detection data based on the pre-trained neural network model to obtain fused features between image data and detection data includes: performing feature fusion processing on the image features and the detection data to obtain the fused features; and performing text description processing on the features to obtain the refined state judgment result corresponding to the features.

4. The method according to claim 2, characterized in that, The pre-trained neural network model includes a visual encoding module and a text encoding module. The step of extracting features from the image data and the detection data based on the pre-trained neural network model to obtain image features corresponding to the image data and text features corresponding to the detection data includes: inputting the image data into the visual encoding module of the neural network model to obtain the image features; and inputting the detection data into the text encoding module of the neural network to obtain the text features.

5. The method according to claim 1, characterized in that, The step of obtaining the bottom blowing control strategy for the LF furnace based on the fusion characteristics, wherein the control strategy is used to describe the steps of setting the bottom blowing data, includes: when the fusion characteristics are in the target refining state stage and the current refining target of the target LF furnace has not been completed, determining that the bottom blowing control setting of the target LF furnace remains unchanged; when the fusion characteristics are in the target refining state stage and the current refining target of the target LF furnace has been completed, determining that the bottom blowing control setting of the target LF furnace is the value of the next stage; wherein the stage is: breaking the top, uniform temperature sampling, adding refining slag, heating and raising the temperature, alloy fine-tuning and soft blowing.

6. The method according to claim 1, characterized in that, After the step of obtaining the LF furnace bottom blowing control strategy based on the fusion characteristics, and the control strategy being used to describe the setting of bottom blowing data, the method further includes: based on the LF furnace bottom blowing control strategy, the control strategy includes refining targets for the LF furnace; based on the LF furnace bottom blowing control strategy, the control strategy takes into account the production and operation status of the LF furnace permeable bricks.

7. The method according to claim 1, characterized in that, The step of establishing a fuzzy logic controller based on the bottom blowing control strategy, adjusting the flow rate and pressure of the bottom blowing gas, and achieving precise control of the bottom blowing of the LF furnace includes: Determine the specific range of flow rate based on the flow rate data of the bottom-blown gas; Determine the specific pressure range based on the pressure data of the bottom-blown gas; Constructing a fuzzy control rule base includes: constructing a fuzzy control rule base based on the specific range of the flow rate and the specific range of the pressure, combined with the requirements of the LF refining process; The acquired LF furnace bottom blowing gas pressure and flow data are fuzzified, including by using one of Gaussian membership function, trapezoidal membership function or triangular membership function to convert the LF furnace bottom blowing gas pressure and flow data into corresponding fuzzy sets; The fuzzy quantities of the LF furnace bottom blowing pressure and flow rate values ​​that need to be adjusted are obtained through fuzzy inference, including: based on the fuzzy control logic rule base and the fuzzy set of bottom blowing gas flow rate and pressure obtained through fuzzification processing, fuzzy inference is performed using the Sugenno inference method to obtain the fuzzy quantities of the LF furnace bottom blowing gas pressure and flow rate values ​​that need to be adjusted; the adjustment fuzzy quantities obtained by fuzzy inference are defuzzified, including: using either the maximum membership degree method or the weighted average division, the adjustment fuzzy values ​​obtained by fuzzy inference are converted into precise flow control values ​​and pressure control values.

8. The method according to claim 1, characterized in that, The step of acquiring sensing data about the target LF furnace includes: controlling the acquisition device to execute a preset acquisition task, wherein the acquisition device is equipped with a variety of sensors, including an image sensor, a pressure sensor, and a flow sensor; acquiring image data of the refining liquid level of the LF furnace based on the image sensor, and acquiring pressure data based on the pressure sensor and flow data based on the flow sensor.

9. An intelligent bottom-blowing stirring control device for an LF furnace, characterized in that, include: The data acquisition module is used to acquire sensing data of the LF refining state to be controlled, including LF furnace molten steel level image data and LF furnace bottom blowing gas detection data; the data feature processing module performs feature extraction processing on the image data and detection data based on a pre-trained neural network model to obtain fused features between the image data and detection data; the control decision module obtains the LF furnace bottom blowing control strategy based on the fused features, the control strategy being used to describe the setting of the bottom blowing data; the control execution module establishes a fuzzy logic controller based on the bottom blowing control strategy, adjusts the flow rate and pressure of the bottom blowing gas, and achieves precise control of the LF furnace bottom blowing.

10. A storage medium for intelligent bottom-blowing stirring control in an LF furnace, characterized in that, include: The program code or instructions are executed by a computer to implement the method described in any one of claims 1 to 8.

Images