Intelligent prediction method and system for residual element content of converter smelting endpoint

Abstract

The application relates to the technical field of smelting element prediction, and discloses an intelligent prediction method and system for residual element content at a converter smelting endpoint. The method comprises the following steps: integrating first smelting data and second smelting data based on heat information to generate a historical sample set; screening sample data belonging to a first threshold interval in a sub-sample set, and combining all sample data into a standard sample set; extracting key parameters in the standard sample set, calculating a first content of residual elements at the converter endpoint based on a preliminary content and a yield, setting a label value for the standard sample set based on the first content to generate a label sample set; summarizing all label sample sets based on the heat information to generate a training data set, constructing a residual element prediction model at the converter endpoint based on the training data set, predicting and outputting residual element content in molten steel at the converter endpoint, and adjusting a production process based on the residual element content. The application improves the efficiency and accuracy of residual element content prediction.

Description

Technical Field

[0001] This application relates to the field of metallurgical element prediction technology, and in particular to an intelligent prediction method and system for the residual element content at the end of converter smelting. Background Technology

[0002] Converter steelmaking is currently the most widely used metallurgical production vessel and method for molten iron steelmaking. The endpoint of converter smelting aims to obtain the endpoint temperature, C, O, P content, and other residual alloying elements suitable for the steel grade. Obtaining the converter endpoint composition parameters in advance allows for understanding the degree of oxidation and alloying content of the molten steel, enabling the pre-configuration of deoxidizers and ferroalloys needed for the deoxidation and alloying stage at the converter tapping stage. This effectively improves the pace of converter smelting, enhances the quality of the molten steel after tapping, and reduces the adjustment pressure during subsequent ladle refining.

[0003] To predict the final state of molten steel at the converter smelting endpoint, numerous converter smelting models have been established. Predicting the residual elements in the molten steel at the endpoint is a crucial component of these models. However, due to the complex physicochemical reactions involving steel, slag, and oxygen within the converter, traditional mechanistic models are ineffective in predicting residual elements at the endpoint, exhibiting significant bias and volatility. With the rise of artificial intelligence, however, a large number of machine learning and deep learning algorithms have been used to build converter endpoint prediction models.

[0004] A similar prior art is disclosed in Chinese patent application CN102392095A, which discloses a method and system for predicting the endpoint of converter steelmaking. The method includes: collecting current production parameter information and current flame information at the converter opening as independent variables; creating a neural network and training the neural network using training samples composed of the independent variables; and using the neural network obtained after training termination to predict the endpoint of converter steelmaking based on the independent variables, obtaining predictions for endpoint time, carbon content, and molten steel temperature. By adopting the technical solution provided by this invention, and by acquiring converter production information and converter opening flame information in real time, the endpoint time, molten steel temperature, and carbon content of steelmaking can be analyzed accurately online in real time, thereby achieving precise control of the process, improving the automation level and production efficiency of converter steelmaking, and reducing costs. Another Chinese patent application, CN113255102A, discloses a method and apparatus for predicting the carbon content and temperature of molten steel at the converter endpoint. The method includes: treating each converter smelting process as a case and describing the process; the case description includes a case feature description and a solution description, where case features include single-value type influencing factors and time-series type influencing factors; the solution includes the carbon content and temperature of the molten steel at the converter endpoint; using the current converter smelting process as the problem case and historical converter smelting processes as historical cases, retrieving similar cases from the historical cases based on a case reasoning algorithm; and obtaining the predicted values ​​of the carbon content and temperature of the molten steel at the converter endpoint for the problem case through case reuse based on the retrieved similar cases and their corresponding carbon content and temperature values. This invention's method has higher prediction accuracy than existing prediction models and can meet the requirements of converter on-site production.

[0005] The shortcomings of existing technologies are mainly reflected in their focus on predicting the carbon content and temperature of molten steel at the end of the converter process. However, in the actual converter steelmaking process, the content of residual elements has a crucial impact on the quality of steel. It is necessary to predict the content of residual elements in order to more comprehensively assess the quality of molten steel at the end of the steelmaking process. Summary of the Invention

[0006] This application provides an intelligent prediction method and system for the residual element content at the end of converter smelting. By making full use of the data after alloying of the steel after converter tapping and the relevant data in the LF refining process, the residual element content at the end of the converter is deduced in reverse, thereby achieving low-cost and high-efficiency prediction of residual elements at the end of converter smelting, and improving the efficiency and accuracy of intelligent prediction of residual element content at the end of converter smelting.

[0007] In a first aspect, this application provides an intelligent prediction method for the residual element content at the end point of converter smelting, the intelligent prediction method for the residual element content at the end point of converter smelting includes:

[0008] First smelting data of the converter smelting process is obtained from the secondary system and MES system, and second smelting data of the LF refining process is obtained. Based on the furnace information, the first smelting data and the second smelting data are integrated to generate a historical sample set.

[0009] The historical sample set is segmented based on steel grade category to generate multiple sub-sample sets. The feature mean and standard deviation corresponding to any steel grade category are calculated based on the sub-sample sets. A first threshold interval is set based on the feature mean and standard deviation. Sample data belonging to the first threshold interval are selected from the sub-sample sets. All the sample data are combined into a standard sample set.

[0010] Key parameters are extracted from the standard sample set. Based on the key parameters, the preliminary content of element C and the yield of residual elements at the converter smelting endpoint are calculated. Based on the preliminary content and the yield, the first content of residual elements at the converter endpoint is calculated. Based on the first content, a label value is set for the standard sample set to generate a label sample set.

[0011] Based on the furnace information, all the labeled sample sets are aggregated to generate a training dataset. A converter endpoint residual element prediction model is constructed based on the training dataset. The converter endpoint residual element prediction model predicts and outputs the residual element content in the molten steel at the converter endpoint. The production process is adjusted based on the residual element content.

[0012] In conjunction with the first aspect, the generation of the historical sample set includes:

[0013] The first smelting data includes the weight data of the steel grade and alloy materials, the alloy batch, and the first component detection data of the alloy materials; the second smelting data includes the weight data of molten steel and the second component detection data of molten steel.

[0014] The first smelting data and the second smelting data are converted into a unified data format. The furnace information is used as the primary key. The first smelting data and the second smelting data are matched and summarized to form sample data. The sample data contains all the characteristics of the converter smelting process and the relevant characteristics of the LF refining process. The second component detection data is set as the component label of the sample data.

[0015] Based on the component labels, the sample data corresponding to all the furnace information are aggregated to generate the historical sample set.

[0016] In conjunction with the first aspect, the extraction of key parameters from the standard sample set includes:

[0017] In the standard sample set, the weight of molten steel in any heat of LF refining process is set as the first data, the weights of multiple alloys added in any heat of converter smelting process are summarized to generate a first column vector, the composition detection data of any alloy batch in any heat of converter smelting process are summarized to generate a first matrix, and the detection results of the first molten steel sampling in any heat of LF refining process are summarized to generate a second column vector.

[0018] The first data, the first column vector, the first matrix, and the second column vector are all set as the key parameters.

[0019] In conjunction with the first aspect, the construction of the converter endpoint residual element prediction model based on the training dataset includes:

[0020] The first content of the residual elements at the converter endpoint is set as the target variable. A correlation analysis algorithm is used to calculate the correlation coefficient between any feature data in the training dataset and the target variable. Feature data with a correlation coefficient greater than a first preset value is set as associated feature data.

[0021] A prediction model is constructed using an LSTM neural network. All the associated feature data and the target variable are divided into a training set, a test set, and a validation set. The prediction model is trained using the training set, and the model weights and bias parameters of the prediction model are adjusted based on the number of iterations and gradient descent.

[0022] Set a callback function, save the model weights corresponding to the validation set based on different iteration numbers, use the test set to evaluate the prediction model corresponding to any model weight, output the evaluation result, and generate the prediction accuracy of the prediction model based on the evaluation result;

[0023] The prediction model with the highest prediction accuracy is set as the prediction model for residual elements at the converter endpoint.

[0024] In conjunction with the first aspect, the method of constructing a prediction model using an LSTM neural network includes:

[0025] Set up a three-layer LSTM network, set a preset number of neurons for any LSTM network layer, and add a Dropout layer after any LSTM network layer;

[0026] After any of the LSTM network layers, a fully connected layer is added. The fully connected layer integrates the output results of any of the LSTM network layers, sets the number of neurons in the output layer based on the number of target variables, and combines all the LSTM network layers, the fully connected layer, and the output layer to form the prediction model.

[0027] In conjunction with the first aspect, the residual element prediction model at the converter endpoint predicts and outputs the residual element content in the molten steel at the converter endpoint, including:

[0028] Obtain the furnace information and feature data corresponding to the molten steel at the converter endpoint, set them as the data to be predicted, input the data to be predicted into the converter endpoint residual element prediction model, and output the prediction result based on the label type of the label value;

[0029] The prediction results are converted from the model output format to the residual element content.

[0030] In conjunction with the first aspect, adjusting the production process based on the residual element content includes:

[0031] Based on the tag type, the target range of the residual element content is set, and the residual element content is compared with the target range to determine whether the smelting process of the current furnace information meets the requirements. If it does not meet the requirements, the tag type corresponding to the residual element content that does not meet the target range is extracted and set as the tag to be adjusted.

[0032] The adjustment type of the label to be adjusted is obtained by comparing it with the label value of the standard sample set. The production process is then adjusted based on the adjustment type, wherein the adjustment type includes the mixing ratio of alloy materials and process parameters.

[0033] In conjunction with the first aspect, anomaly detection is performed on the historical sample set, including:

[0034] Normal smelting data is extracted from the historical sample set corresponding to different furnace information. The normal smelting data is used to train an isolated forest model, multiple random trees are established and the isolation degree of each data point is calculated. An isolation score is assigned to any data point based on the isolation degree, and a second threshold is set based on all the isolation scores.

[0035] The remaining smelting data in the historical sample set is input into the isolated forest model to calculate a new isolation score. If the new isolation score is greater than the second threshold, the remaining smelting data is determined to be an abnormal data point.

[0036] The data source of the abnormal data points is obtained, and anomaly analysis results are generated based on the data source. The anomaly analysis results include equipment failure prediction, operation deviation identification, and raw material unevenness detection.

[0037] Secondly, this application provides an intelligent prediction system for the residual element content at the end of converter smelting, the intelligent prediction system for the residual element content at the end of converter smelting includes:

[0038] The data acquisition module is used to acquire the first smelting data of the converter smelting process from the secondary system and the MES system, acquire the second smelting data of the LF refining process, and integrate the first smelting data and the second smelting data based on the furnace information to generate a historical sample set.

[0039] The filtering module is used to segment the historical sample set according to the steel grade category to generate multiple sub-sample sets, calculate the feature mean and standard deviation corresponding to any steel grade category based on the sub-sample sets, set a first threshold interval based on the feature mean and the standard deviation, filter out sample data belonging to the first threshold interval from the sub-sample sets, and combine all the sample data into a standard sample set.

[0040] The label setting module is used to extract key parameters from the standard sample set, calculate the preliminary content of element C and the yield of residual elements at the converter smelting endpoint based on the key parameters, calculate the first content of residual elements at the converter endpoint based on the preliminary content and the yield, set label values ​​for the standard sample set based on the first content, and generate a label sample set.

[0041] The prediction module is used to summarize all the label sample sets according to the furnace information, generate a training dataset, construct a converter endpoint residual element prediction model based on the training dataset, predict and output the residual element content in the molten steel at the converter endpoint, and adjust the production process based on the residual element content.

[0042] The technical solution provided in this application firstly utilizes the steel composition detection results from the LF refining stage and data from the deoxidation and alloying process at the converter tapping stage to reverse-calculate the residual element content at the converter endpoint. This eliminates the need for sampling and testing at the converter endpoint, significantly reducing production costs. Through data analysis and calculation, label values ​​for residual elements at the converter endpoint are automatically generated, solving the problem of traditional methods being unable to use supervised learning methods due to a lack of labeled data, thus providing a foundation for establishing an accurate prediction model. Secondly, a three-layer LSTM network structure is employed, effectively capturing the time dependence of various feature data during the converter smelting process, improving the model's ability to model complex metallurgical processes. Adding a Dropout layer after each LSTM layer effectively prevents overfitting and improves the model's generalization ability. Correlation analysis is used to select feature data strongly correlated with the target variable, reducing model complexity and improving training efficiency and prediction accuracy. Finally, an isolation forest model is used to detect anomalies in the historical sample set, effectively identifying data points that do not conform to the normal production process and proactively identifying potential equipment failures, operational deviations, and uneven raw material distribution. Based on prediction results and anomaly analysis, production process parameters are dynamically adjusted to ensure production stability and product quality. By predicting the residual element content at the converter endpoint in advance, the deoxidizers and ferroalloys required for the deoxidation and alloying stage of converter steelmaking can be configured ahead of time, optimizing the production process and improving production efficiency.

[0043] This application, based on trace analysis of element source-loss-destination in the metallurgical process, uses the weighing and sampling analysis results of molten steel in LF refining, combined with the weight and composition of alloy material weighed during the alloying stage at the converter tapping stage, and the alloy element yield converted to C element ranges. It presents a low-cost method for obtaining the residual element content at the converter smelting endpoint without sampling. This method does not rely on sampling and analysis of molten steel at the converter endpoint; the missing residual element content labels at the converter endpoint can be obtained through pure data analysis, processing, and calculation. This smelting mode, which does not perform elemental analysis on molten steel samples at the converter smelting endpoint, can quickly and cost-effectively obtain a training set for establishing a predictive model for residual elements at the converter endpoint. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of an embodiment of the intelligent prediction method for residual element content at the end point of converter smelting in this application.

[0046] Figure 2 This is a schematic diagram of one embodiment of the label value setting process in this application.

[0047] Figure 3 This is a schematic diagram of one embodiment of the production process adjustment process in this application.

[0048] Figure 4 This is a schematic diagram of an embodiment of an intelligent prediction system for the residual element content at the end of converter smelting in this application. Detailed Implementation

[0049] This application provides an intelligent prediction method and system for the residual element content at the end point of converter smelting. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0050] For models built using artificial intelligence (AI) technology, their predictive performance and accuracy primarily depend on two factors: the sample set used for model training and the AI ​​algorithm employed. The sample set used for training is the most crucial, determining the upper limit of the model's accuracy, while the algorithm merely approximates this limit. Currently, AI models applied in the metallurgical industry mainly employ supervised learning methods because they primarily predict various quantitative data. For the converter smelting model dataset, the data obtained during the smelting process form the feature columns of the dataset, while the residual element data at the converter endpoint serves as the labels for each row of samples in the dataset. Labeling the residual elements at the converter endpoint relies mainly on sampling and detection at the converter smelting endpoint, a process requiring significant investment in samplers and labor.

[0051] This application fully utilizes data from the converter steelmaking process after alloying, employing the composition analysis results of the first sample of molten steel taken from the refining ladle and the analysis results of various alloy additions and alloy compositions during the converter deoxidation and alloying process. The residual element content of the final molten steel in the converter is deduced from the composition analysis results of the LF refining samples. This method can supplement the labeling data in the complete converter model data sample set. Furthermore, this supplementation method does not rely on manual sampling and composition analysis for labeling, significantly reducing material and labor costs while obtaining relatively accurate labels for residual elements at the converter endpoint.

[0052] The basic principle of this application is the mass conservation of alloying elements in molten steel. By analyzing the characteristics of the smelting process, a source-loss-fate trace analysis model for alloying elements in molten steel is established. After the converter reaches the end of blowing, the residual element content in the molten steel stabilizes. When the converter is emptied and tapped, deoxidation and alloying operations are generally performed. At this time, the alloying elements in the molten steel are replenished by alloying elements brought in from external alloys. However, not all of these alloying elements enter the molten steel; some are lost through oxidation, forming inclusions or entering the slag. After tapping and deoxidation and alloying are completed, the materials containing alloying elements have been added, and the element content in the molten steel in the ladle stabilizes again. Due to the effect of bottom blowing argon gas in the ladle, the components are continuously homogenized. At the LF ladle refining process, the molten steel is weighed and its temperature is measured before any materials are added. At this time, the total mass of various elements in the molten steel can be calculated. At the same time, based on the mass and composition detection results of various alloys added during the deoxidation and alloying stage, the theoretical total mass of various alloying elements brought into the molten steel by the alloying materials can be calculated. First, based on the above data, the carbon content at the converter endpoint is preliminarily calculated, the yield is calculated, and the losses of various alloying elements carried into the molten steel by the alloying materials are determined to obtain the actual mass of each element carried into the molten steel. Finally, according to the law of conservation of element mass, the mass of each element in the molten steel at the converter endpoint is obtained by subtracting the mass of elements carried into the molten steel by the alloying materials from the total mass of elements sampled at LF. Combined with the mass of the molten steel, the content of residual elements at the converter endpoint can then be calculated.

[0053] This method eliminates the need to modify the original low-cost converter tapping model that does not involve sampling molten steel. It utilizes data from converter tapping and LF refining to process and compute information about residual elements in the molten steel at the converter endpoint, thus creating a complete sample dataset for supervised learning of residual elements at the converter endpoint. This serves as a training set for establishing machine learning models to predict residual elements at the converter endpoint.

[0054] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent prediction method for residual element content at the end point of converter smelting in this application includes:

[0055] Step S101: Obtain the first smelting data of the converter smelting process from the secondary system and the MES system, and obtain the second smelting data of the LF refining process. Based on the furnace information, integrate the first smelting data and the second smelting data to generate a historical sample set.

[0056] It is understood that the executing entity of this application can be an intelligent prediction device for the residual element content at the end of converter smelting, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiment uses a server as an example for illustration.

[0057] Specifically, the first smelting data includes the steel grade being smelted, the weight data and batch numbers of various alloy materials added during the converter smelting process, and the composition analysis data of all alloy materials entering the furnace for the corresponding batch. The second smelting data includes the weight data of the molten steel in the ladle, and the data from the first composition analysis of the molten steel. The converter smelting data and LF refining data for the same heat are matched and summarized to form a sample data, ultimately forming a historical sample set.

[0058] Step S102: Divide the historical sample set based on steel grade category to generate multiple sub-sample sets. Calculate the feature mean and standard deviation corresponding to any steel grade category based on the sub-sample sets. Set a first threshold interval based on the feature mean and standard deviation. Select sample data belonging to the first threshold interval from the sub-sample sets and combine all sample data into a standard sample set.

[0059] Specifically, in the historical sample set, the characteristics of each acquired sample are screened and judged. First, samples with missing features are deleted. Then, the historical sample set is divided into multiple sub-sample sets according to steel type, with each sub-sample set corresponding to a different steel type. (Feature mean) The calculation formula is: Where N is the number of subsets. Let be the feature of the i-th sample in the subsample set. Standard deviation The calculation formula is: The first threshold interval can be set to [ -3 ， +3 ] Delete sample data outside the first threshold interval, and keep sample data within the first threshold interval to form a standard sample set.

[0060] Step S103: Extract key parameters from the standard sample set, calculate the preliminary content of element C and the yield of residual elements at the converter smelting endpoint based on the key parameters, calculate the first content of residual elements at the converter endpoint based on the preliminary content and the yield, set label values ​​for the standard sample set based on the first content, and generate a label sample set.

[0061] Specifically, Figure 2 This is a flowchart for setting label values. Key parameters refer to the parameter values ​​used to predict residual elements. The yield of residual elements (C, Si, Mn, Cr, etc.) is calculated based on the residual C content at the converter endpoint. Among these, the key parameters include the weight of molten steel measured during a single heat of LF refining. m iron (tons), the weight of various alloys added during the tapping process of a single heat in the converter. m alloy (kg), batch alloy composition test data [ %e ] alloy Results of element content analysis in the first molten steel sample taken during the LF refining process. %e ] LF When e=C, the final C content of the converter smelting is calculated by subtracting the total C content introduced by all alloys from the C content in the LF molten steel. This is the initial content of element C. The calculation formula is as follows: ,in, This is the transpose of the column vector of C content of the i alloys added to the molten steel during the alloying stage. The yields of elements e=C, Si, Mn, and Cr are calculated using the following formulas:

[0062]

[0063]

[0064]

[0065]

[0066] in, , , and These represent the yields of the residual elements C, Si, Mn, and Cr, respectively. Losses of other elements upon entering the molten steel are negligible, and their yields are denoted as 1. There are a total of j elements; therefore, the element yields in each alloy are delimited as a column vector. η e for: The first content of residual elements at the converter endpoint The calculation formula is: Among them, the calculated residual element content at the converter endpoint [ %e ] f as one j A row and column vector, where each row represents the content of each element in the molten steel at the end of the converter cooling process. The content of j residual elements at the converter endpoint of a single furnace was calculated. The same calculation process was performed on each sample in the sample set to quickly obtain the label values ​​of the residual elements at the converter endpoint of each sample, thus generating a label sample set.

[0067] Step S104: Summarize all tag sample sets based on furnace information, generate a training dataset, construct a converter endpoint residual element prediction model based on the training dataset, predict and output the residual element content in the molten steel at the converter endpoint based on the residual element content, and adjust the production process based on the residual element content.

[0068] Specifically, a residual element prediction model is constructed based on the training dataset. This model can be trained using machine learning algorithms (such as LSTM). The model predicts the residual element content in the molten steel at the converter endpoint. Based on the model's predicted residual element content, the converter smelting process is adjusted to optimize steel quality.

[0069] In one specific embodiment, generating a historical sample set includes:

[0070] (1) The first smelting data includes the weight data of the smelting steel grade and alloy materials, the alloy batch, and the first component test data of the alloy materials. The second smelting data includes the weight data of molten steel and the second component test data of molten steel.

[0071] (2) Convert the first smelting data and the second smelting data into a unified data format, use the furnace information as the primary key, match the first smelting data and the second smelting data and summarize them to form sample data. The sample data contains all the features of the converter smelting process and the relevant features of the LF refining process. Set the second component detection data as the component label of the sample data.

[0072] (3) Based on the component labels, the sample data corresponding to all furnace information are summarized to generate a historical sample set.

[0073] Specifically, the steel grade being smelted refers to the target steel grade being smelted in this heat. The weight data of the alloy materials refers to the recorded weight of various alloy materials added during the converter tapping process. The alloy batch refers to the recorded batch information of the added alloy materials, used for tracing the alloy composition testing data. The first component testing data of the alloy materials refers to the recorded composition testing results of each batch of alloy materials, including the content of various elements. The molten steel weight data refers to the recorded weight of the molten steel in the ladle during the LF refining process. The second component testing data of the molten steel refers to the recorded composition testing results of the first molten steel sample taken during the LF refining process, including the content of various elements.

[0074] The first and second smelting data were converted into a unified data format to ensure data consistency and operability. Using furnace number information as the primary key, the first and second smelting data were matched. The specific steps are as follows: Furnace number information, alloy material weight, alloy batch, and alloy composition detection data were extracted from the first smelting data; furnace number information, molten steel weight, and molten steel composition detection data were extracted from the second smelting data; based on the furnace number information, the converter smelting data and LF refining data were matched one-to-one to form complete sample data. Each sample data entry contains all characteristics of the converter smelting process (such as steel grade, alloy material weight, alloy batch, alloy composition, etc.) and relevant characteristics of the LF refining process (such as molten steel weight, molten steel composition, etc.). The composition detection data of the first molten steel sample taken during the LF refining process was set as the composition label for this sample data. The sample data corresponding to all furnace number information were summarized to generate a historical sample set.

[0075] In one specific embodiment, key parameters are extracted from a standard sample set, including:

[0076] (1) In the standard sample set, the weight of molten steel in any heat of LF refining process is set as the first data, the weight of multiple alloys added in any heat of converter smelting process is summarized to generate the first column vector, the composition detection data of any alloy batch in any heat of converter smelting process is summarized to generate the first matrix, and the detection results of the first molten steel sampling in any heat of LF refining process are summarized to generate the second column vector.

[0077] (2) Set the first data, the first column vector, the first matrix and the second column vector as key parameters.

[0078] Specifically, the key parameters are represented as the first data. m iron The weights of the i alloys are summarized to generate the first column vector. i alloys were added during the converter tapping process, and each alloy was tested. j kind e element( e = C , Si , Mn , Cr ...the content of j elements in total, and the weight of various alloys added [ %e ] alloy Let it be the first matrix Test results %e ] LF Let it be the second column vector .

[0079] In one specific embodiment, a converter endpoint residual element prediction model is constructed based on a training dataset, including:

[0080] (1) Set the first content of residual elements at the converter endpoint as the target variable, use the correlation analysis algorithm to calculate the correlation coefficient between any feature data in the training dataset and the target variable, and set the feature data with a correlation coefficient greater than the first preset value as the associated feature data.

[0081] (2) Use LSTM neural network to build a prediction model, divide all related feature data and target variables into training set, test set and validation set, use training set to train prediction model, and adjust model weight and bias parameters of prediction model based on iteration number and gradient descent.

[0082] (3) Set a callback function, save the model weights corresponding to the validation set based on different iterations, use the test set to evaluate the prediction model corresponding to any model weight, output the evaluation results, and generate the prediction accuracy of the prediction model based on the evaluation results.

[0083] (4) Set the prediction model with the highest prediction accuracy as the converter end point residual element prediction model.

[0084] Specifically, the target variable serves as the label value of the sample data. Correlation analysis algorithms (such as Pearson correlation coefficient) can be used to evaluate the degree of correlation between two variables in the training dataset and filter out the feature data that is strongly correlated with the target variable, which are the associated feature data.

[0085] Long Short-Term Memory (LSTM) networks are a special type of recurrent neural network (RNN) capable of effectively handling long-term dependencies in time-series data. In the converter smelting process, there are temporal dependencies between various feature data, and LSTM can effectively capture these dependencies. The associated feature data and target variable are partitioned according to a ratio of 70% training set, 15% validation set, and 15% test set. Gradient descent is an optimization algorithm used to minimize a loss function. During training, the model's weights and biases are updated by calculating the gradient of the loss function with respect to the model parameters. The specific steps are as follows: Initialize the model's weights and biases; in each iteration, calculate the loss function between the model's predicted and actual values; update the model's weights and biases using the gradient descent algorithm to minimize the loss function; until the preset number of iterations is reached or the loss function converges.

[0086] A callback function is a mechanism used to monitor model performance during training. In this invention, the callback function is used to save the model weights corresponding to the validation set at different iteration numbers. The root mean square error is used as the evaluation result. The prediction model corresponding to each saved model weight is evaluated using the test set; the model weights refer to the trainable parameters in the model. The prediction model with the highest prediction accuracy is selected as the final prediction model for the residual elements at the converter endpoint.

[0087] In one specific embodiment, a prediction model is constructed using an LSTM neural network, including:

[0088] (1) Set up a three-layer LSTM network, set a preset number of neuron units for any LSTM network layer, and add a Dropout layer after any LSTM network layer.

[0089] (2) After any LSTM network layer, add a fully connected layer. The fully connected layer integrates the output results of any LSTM network layer. The number of neurons in the output layer is set based on the number of target variables. All LSTM network layers, fully connected layers and output layers are combined to form a prediction model.

[0090] Specifically, LSTM network layers are used to capture the temporal dependencies of various feature data during the converter smelting process. Neurons are the basic building blocks of the LSTM network, each containing an input gate, a forget gate, and an output gate to control the flow of information. Dropout is a regularization technique used to prevent overfitting in neural networks. During training, the Dropout layer randomly discards a portion of the neurons' outputs, thereby improving the model's generalization ability.

[0091] Fully connected layers are used to integrate the outputs of the LSTM network layers and map the results to the dimension of the target variable. The output of the fully connected layer is the model's prediction. The target variable is the content of residual elements at the converter endpoint; therefore, the number of neurons in the output layer equals the number of types of residual elements. All LSTM network layers, Dropout layers, fully connected layers, and the output layer are combined to form the prediction model.

[0092] In one specific embodiment, the converter endpoint residual element prediction model predicts and outputs the residual element content in the molten steel at the converter endpoint, including:

[0093] (1) Obtain the furnace information and characteristic data corresponding to the molten steel at the end of the converter, set them as the data to be predicted, input the data to be predicted into the residual element prediction model at the end of the converter, and output the prediction results based on the label type of the label value.

[0094] (2) Convert the prediction results from the model output format to the residual element content.

[0095] Specifically, the data to be predicted refers to the heat information and characteristic data corresponding to the final molten steel in the converter. This data is used to predict the content of residual elements at the converter endpoint after the model training is completed. The label type of the label value refers to the format of the model output. The format of the model output is the content of residual elements, so the label type is continuous numerical.

[0096] The model output is the content of residual elements, usually a vector, where each element corresponds to a predicted value of a residual element. The model output prediction results are converted from the model's internal format (such as tensors or arrays) into the actual residual element content.

[0097] In one specific embodiment, adjusting the production process based on the residual element content includes:

[0098] (1) Set the target range of residual element content based on the label type, compare the residual element content with the target range, and determine whether the smelting process of the current furnace information meets the requirements. If it does not meet the requirements, extract the label type corresponding to the residual element content that does not meet the target range and set it as the label to be adjusted.

[0099] (2) Compare the label to be adjusted with the label value of the standard sample set to obtain the adjustment type of the label to be adjusted, and adjust the production process based on the adjustment type. The adjustment type includes the mixing ratio of alloy materials and process parameters.

[0100] Specifically, Figure 3Adjust the flow chart for the production process. The target range refers to the ideal range of residual element content at the converter endpoint during production. The label type refers to the type of residual element, such as C, Si, Mn, Cr, etc. The labels to be adjusted refer to the label types corresponding to residual element contents that do not conform to the target range. The adjustment type refers to the production process parameters that need to be adjusted, including the mixing ratio of alloy materials and process parameters. For example, if analysis of a standard sample set reveals that the C element content is too high, it is usually necessary to reduce the amount of low-carbon ferromanganese added or adjust process parameters such as the blowing time.

[0101] In one specific embodiment, anomaly detection is performed on the historical sample set, including:

[0102] (1) Extract normal smelting data from the historical sample set corresponding to different furnace information, train the isolated forest model using normal smelting data, establish multiple random trees and calculate the isolation degree of each data point, assign an isolation score to any data point based on the isolation degree, and set a second threshold based on all isolation scores.

[0103] (2) Input the remaining smelting data in the historical sample set into the isolated forest model and calculate the new isolated score. If the new isolated score is greater than the second threshold, the remaining smelting data is determined to be abnormal data points.

[0104] (3) Obtain the data source of abnormal data points and generate abnormal analysis results based on the data source. The abnormal analysis results include equipment failure prediction, operation deviation identification, and raw material unevenness detection.

[0105] Specifically, the historical sample set refers to converter smelting data from multiple heats, including various characteristics (such as steel grade, alloy addition amount, alloy composition, etc.) and labels (such as residual element content). Normal smelting data refers to data points in the historical sample set that conform to the normal production process; this data is used to train the Isolation Forest model. For example, normal data points can be selected through statistical analysis; normal data points should have stable production parameters and residual element content that meets quality standards. Isolation Forest is a tree-based anomaly detection algorithm that isolates data points by randomly partitioning the feature space, thereby detecting anomalies. The Isolation Forest model builds multiple random trees by randomly partitioning the feature space. An isolation score is a score assigned to each data point by the Isolation Forest model, measuring the degree of isolation of that data point; the higher the score, the more likely the data point is to be an anomaly. The second threshold is a threshold set based on the isolation score, used to determine whether a data point is an anomaly.

[0106] If the isolated score of a data point in the remaining smelting data is 0.98, exceeding the second threshold of 0.95, then that data point is identified as an outlier. Anomaly analysis results refer to the conclusions drawn after analyzing anomalous data points, including equipment failure prediction, operational deviation identification, and raw material inhomogeneity detection. For example, if an anomalous data point is characterized by an abnormally high alloy addition, it may be due to equipment failure or operational deviation; if it is characterized by large fluctuations in residual element content, it may be due to raw material inhomogeneity.

[0107] The above describes the intelligent prediction method for residual element content at the endpoint of converter smelting in the embodiments of this application. The following describes the intelligent prediction system for residual element content at the endpoint of converter smelting in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 4 One embodiment of the intelligent prediction system for residual element content at the end point of converter smelting in this application includes:

[0108] The data acquisition module 201 is used to acquire the first smelting data of the converter smelting process from the secondary system and the MES system, acquire the second smelting data of the LF refining process, and integrate the first smelting data and the second smelting data based on the furnace information to generate a historical sample set.

[0109] The filtering module 202 is used to segment the historical sample set according to the steel grade category, generate multiple sub-sample sets, calculate the feature mean and standard deviation corresponding to any steel grade category based on the sub-sample sets, set a first threshold interval based on the feature mean and standard deviation, filter out sample data belonging to the first threshold interval from the sub-sample sets, and combine all sample data into a standard sample set.

[0110] The label setting module 203 is used to extract key parameters from the standard sample set, calculate the preliminary content of element C and the yield of residual elements at the converter smelting endpoint based on the key parameters, calculate the first content of residual elements at the converter endpoint based on the preliminary content and the yield, set label values ​​for the standard sample set based on the first content, and generate a label sample set.

[0111] The prediction module 204 is used to summarize all the label sample sets according to the furnace information, generate a training dataset, build a converter end point residual element prediction model based on the training dataset, predict the residual element content in the molten steel at the converter end point, and adjust the production process based on the residual element content.

[0112] Through the collaborative efforts of the aforementioned components, firstly, by utilizing the steel composition detection results from the LF refining stage and data from the converter tapping deoxidation and alloying process, the residual element content at the converter endpoint is calculated in reverse, eliminating the need for sampling and testing at the converter endpoint and significantly reducing production costs. Secondly, through data analysis and calculation, label values ​​for residual elements at the converter endpoint are automatically generated, solving the problem of traditional methods being unable to use supervised learning methods due to a lack of labeled data, thus providing a foundation for establishing an accurate prediction model. Thirdly, a three-layer LSTM network structure is employed, effectively capturing the time dependencies of various feature data during the converter smelting process, improving the model's ability to model complex metallurgical processes. Adding a Dropout layer after each LSTM layer effectively prevents overfitting and improves the model's generalization ability. Correlation analysis is used to filter feature data strongly correlated with the target variable, reducing model complexity and improving training efficiency and prediction accuracy. Finally, an isolation forest model is used to detect anomalies in the historical sample set, effectively identifying data points that do not conform to the normal production process and proactively identifying potential equipment failures, operational deviations, and uneven raw material distribution. Based on prediction results and anomaly analysis, production process parameters are dynamically adjusted to ensure production stability and product quality. By predicting the residual element content at the converter endpoint in advance, the deoxidizers and ferroalloys required for the deoxidation and alloying stage of converter steelmaking can be configured ahead of time, optimizing the production process and improving production efficiency.

[0113] This application, based on trace analysis of element source-loss-destination in the metallurgical process, uses the weighing and sampling analysis results of molten steel in LF refining, combined with the weight and composition of alloy material weighed during the alloying stage at the converter tapping stage, and the alloy element yield converted to C element ranges. It presents a low-cost method for obtaining the residual element content at the converter smelting endpoint without sampling. This method does not rely on sampling and analysis of molten steel at the converter endpoint; the missing residual element content labels at the converter endpoint can be obtained through pure data analysis, processing, and calculation. This smelting mode, which does not perform elemental analysis on molten steel samples at the converter smelting endpoint, can quickly and cost-effectively obtain a training set for establishing a predictive model for residual elements at the converter endpoint.

[0114] Case 1

[0115] A steel plant uses a 100t converter for steelmaking. Since the steel produced by the converter is mostly low-carbon steel, the plant adopts a one-blow-through mode in the converter smelting stage, transferring the task of steel deoxidation and steel composition adjustment to the LF refining department. In order to reduce the cost of the converter smelting stage, when the converter smelting reaches the end, the steel is not sampled for elemental analysis, and the steel is directly deoxidized and alloyed.

[0116] To enhance control over steel quality during the converter smelting stage, the plant needs to establish a converter control model, including a predictive model for residual elements at the converter endpoint. This model is built using the LSTM supervised learning algorithm. However, its converter sample dataset lacks complete sample labeling because it doesn't analyze the residual element content in the final molten steel sample. Therefore, the method described in this invention is used to calculate and complete the residual element label values ​​at the endpoint.

[0117] Taking one heat as an example, the data processing process is explained as follows: First, the features of the datasets of the converter smelting and LF refining processes are extracted. The extraction results are shown in Table 1. The alloy composition matched by the alloy batch number is shown in Table 2. The alloy addition mass is shown in Table 3. The composition content of the first sample of LF molten steel is shown in Table 4.

[0118] Table 1. Characteristic parameters of converter and LF refining extracted from a single heat.

[0119]

[0120] Table 2. Composition test results of the alloys used in this furnace.

[0121]

[0122] Table 3 Alloy addition amount for this batch

[0123]

[0124] Table 4. Compositional analysis results of the first LF steel sample taken from this heat.

[0125]

[0126] The theoretically lossless mass of alloying elements introduced into molten steel by alloying materials is: m alloy-j Substitute the data from the table above:

[0127]

[0128] Right now:

[0129]

[0130] The mass of alloying elements in the molten steel during the first LF sampling was m alloy-j Substitute the data from the table above:

[0131]

[0132] Right now:

[0133]

[0134] Preliminary calculation of C content at the converter endpoint:

[0135]

[0136] Based on the preliminary calculation of C content at the converter endpoint, the recovery rates of each element were calculated as follows:

[0137]

[0138] The residual element content of each element at the converter endpoint was calculated as follows:

[0139]

[0140]

[0141] The endpoint residual element label values ​​of the converter sample dataset were calculated using the same data acquisition, preprocessing, and calculation methods for other heats. This sample set was then used as the training set for the converter endpoint residual element prediction model based on LSTM. The training set performance was validated by predicting the endpoint residual elements of 15 consecutive heats and by sampling molten steel samples. The model prediction results and sampling composition analysis results are shown in the table, where e... p This represents the model prediction value of element e, e a This represents the sampled detection value of element e:

[0142]

[0143] Compare the root mean square error of the above prediction results with the root mean square error of the molten steel sampling results:

[0144]

[0145] The root mean square error of the predictions shows that the prediction model for residual elements at the converter endpoint achieves good prediction accuracy. This indicates that the error between the missing residual element label values ​​at the converter endpoint calculated using the method described in this invention and the true values ​​is very small. Therefore, using this as a training set can enable the trained prediction model to have good prediction accuracy.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0147] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0148] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 this application.

Claims

1. A method for intelligent prediction of residual element content at the end point of converter smelting, characterized in that, The method includes: First smelting data of the converter smelting process is obtained from the secondary system and MES system, and second smelting data of the LF refining process is obtained. Based on the furnace information, the first smelting data and the second smelting data are integrated to generate a historical sample set. The historical sample set is segmented based on steel grade category to generate multiple sub-sample sets. The feature mean and standard deviation corresponding to any steel grade category are calculated based on the sub-sample sets. A first threshold interval is set based on the feature mean and standard deviation. Sample data belonging to the first threshold interval are selected from the sub-sample sets. All the sample data are combined into a standard sample set. Key parameters are extracted from the standard sample set. Based on the key parameters, the preliminary content of element C and the yield of residual elements at the converter smelting endpoint are calculated. Based on the preliminary content and the yield, the first content of residual elements at the converter endpoint is calculated. Based on the first content, a label value is set for the standard sample set to generate a label sample set. Based on the furnace information, all the label sample sets are summarized to generate a training dataset. Based on the training dataset, a converter endpoint residual element prediction model is constructed. The converter endpoint residual element prediction model predicts and outputs the residual element content in the molten steel at the converter endpoint. The production process is adjusted based on the residual element content. The construction of a prediction model for residual elements at the converter endpoint includes: The first content of the residual elements at the converter endpoint is set as the target variable. A correlation analysis algorithm is used to calculate the correlation coefficient between any feature data in the training dataset and the target variable. Feature data with a correlation coefficient greater than a first preset value is set as associated feature data. A prediction model is constructed using an LSTM neural network. All the associated feature data and the target variable are divided into a training set, a test set, and a validation set. The prediction model is trained using the training set, and the model weights and bias parameters of the prediction model are adjusted based on the number of iterations and gradient descent. Set a callback function, save the model weights corresponding to the validation set based on different iteration numbers, use the test set to evaluate the prediction model corresponding to any model weight, output the evaluation result, and generate the prediction accuracy of the prediction model based on the evaluation result; The prediction model with the highest prediction accuracy is set as the prediction model for residual elements at the converter endpoint.

2. The method according to claim 1, characterized in that, The generated historical sample set includes: The first smelting data includes the weight data of the steel grade and alloy materials, the alloy batch, and the first component detection data of the alloy materials; the second smelting data includes the weight data of molten steel and the second component detection data of molten steel. The first smelting data and the second smelting data are converted into a unified data format. The furnace information is used as the primary key. The first smelting data and the second smelting data are matched and summarized to form sample data. The sample data contains all the features of the converter smelting process and the relevant features of the LF refining process. The second component detection data is set as the component label of the sample data. Based on the component labels, the sample data corresponding to all the furnace information are aggregated to generate the historical sample set.

3. The method according to claim 1, characterized in that, The extraction of key parameters from the standard sample set includes: In the standard sample set, the weight of molten steel in any heat of LF refining process is set as the first data, the weights of multiple alloys added in any heat of converter smelting process are summarized to generate a first column vector, the composition detection data of any alloy batch in any heat of converter smelting process are summarized to generate a first matrix, and the detection results of the first molten steel sampling in any heat of LF refining process are summarized to generate a second column vector. The first data, the first column vector, the first matrix, and the second column vector are all set as the key parameters.

4. The method according to claim 1, characterized in that, The method of using an LSTM neural network to build a prediction model includes: Set up a three-layer LSTM network, set a preset number of neurons for any LSTM network layer, and add a Dropout layer after any LSTM network layer; After any of the LSTM network layers, a fully connected layer is added. The fully connected layer integrates the output results of any of the LSTM network layers, sets the number of neurons in the output layer based on the number of target variables, and combines all the LSTM network layers, the fully connected layer, and the output layer to form the prediction model.

5. The method according to claim 1, characterized in that, The converter endpoint residual element prediction model predicts and outputs the residual element content in the molten steel at the converter endpoint, including: Obtain the furnace information and feature data corresponding to the molten steel at the converter endpoint, set them as the data to be predicted, input the data to be predicted into the converter endpoint residual element prediction model, and output the prediction result based on the label type of the label value; The prediction results are converted from the model output format to the residual element content.

6. The method according to claim 1, characterized in that, The process adjustment based on the residual element content includes: Based on the tag type, the target range of the residual element content is set, and the residual element content is compared with the target range to determine whether the smelting process of the current furnace information meets the requirements. If it does not meet the requirements, the tag type corresponding to the residual element content that does not meet the target range is extracted and set as the tag to be adjusted. The adjustment type of the label to be adjusted is obtained by comparing it with the label value of the standard sample set. The production process is then adjusted based on the adjustment type, wherein the adjustment type includes the mixing ratio of alloy materials and process parameters.

7. The method according to claim 1, characterized in that, Anomaly detection is performed on the historical sample set, including: Normal smelting data is extracted from the historical sample set corresponding to different furnace information. The normal smelting data is used to train an isolated forest model, multiple random trees are established and the isolation degree of each data point is calculated. An isolation score is assigned to any data point based on the isolation degree, and a second threshold is set based on all the isolation scores. The remaining smelting data in the historical sample set is input into the isolated forest model to calculate a new isolation score. If the new isolation score is greater than the second threshold, the remaining smelting data is determined to be an abnormal data point. The data source of the abnormal data points is obtained, and anomaly analysis results are generated based on the data source. The anomaly analysis results include equipment failure prediction, operation deviation identification, and raw material unevenness detection.

8. An intelligent prediction system for the residual element content at the end point of converter smelting, used to implement the method as described in any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to acquire the first smelting data of the converter smelting process from the secondary system and the MES system, acquire the second smelting data of the LF refining process, and integrate the first smelting data and the second smelting data based on the furnace information to generate a historical sample set. The filtering module is used to segment the historical sample set according to the steel grade category to generate multiple sub-sample sets, calculate the feature mean and standard deviation corresponding to any steel grade category based on the sub-sample sets, set a first threshold interval based on the feature mean and the standard deviation, filter out sample data belonging to the first threshold interval from the sub-sample sets, and combine all the sample data into a standard sample set. The label setting module is used to extract key parameters from the standard sample set, calculate the preliminary content of element C and the yield of residual elements at the converter smelting endpoint based on the key parameters, calculate the first content of residual elements at the converter endpoint based on the preliminary content and the yield, set label values ​​for the standard sample set based on the first content, and generate a label sample set. The prediction module is used to summarize all the label sample sets according to the furnace information, generate a training dataset, construct a converter endpoint residual element prediction model based on the training dataset, predict and output the residual element content in the molten steel at the converter endpoint, and adjust the production process based on the residual element content.

Images