CN118737313A - A real-time prediction method for molten steel quality based on multi-source data feature fusion - Google Patents

Abstract

The present invention discloses a method for real-time prediction of molten steel quality based on multi-source data feature fusion, and relates to the field of converter steelmaking. It includes constructing a training data set based on feature fusion of multi-source data such as images, audio, flue gas data, raw material information, auxiliary material addition data, oxygen blowing parameters, etc., training the existing XGBoost algorithm, using the trained XGBoost model as a real-time prediction model for molten steel quality, and the output of the model as the predicted value of the molten steel temperature and carbon content at the next moment; transferring the real-time blowing data of a new batch into the real-time prediction model for molten steel quality, and predicting the current real-time carbon content and real-time temperature of the molten steel. Compared with the method of using a single source of data to predict the quality of molten steel in real time, the method of the present invention comprehensively utilizes the common multi-source data of the converter to predict the quality of molten steel in real time during the converter steelmaking blowing process, and has the advantages of strong anti-interference ability, good real-time performance, high precision, and support for converter intelligence.

Description

A real-time prediction method for molten steel quality based on multi-source data feature fusion

Technical Field

The invention relates to the field of converter steelmaking, and in particular to a real-time prediction method for molten steel quality based on multi-source data feature fusion.

Background Art

The output of converter steelmaking accounts for more than 70% of the total steel production, and it is the mainstream steelmaking technology in the world today. Among them, the quality control of molten steel in the converter steelmaking process is directly related to the quality of the tapped molten steel. Due to the high temperature and high smoke concentration in the production environment of the converter steelmaking process, the real-time monitoring of the production process is restricted. At present, the analytical methods of the converter steelmaking process include flue gas analysis technology, furnace mouth flame analysis technology, manual experience judgment, and auxiliary gun measurement technology. Among them, the auxiliary gun measurement technology is the most widely used. However, this method has defects such as destroying production continuity, increasing steelmaking costs, and single-point limitations of sampling points, which potentially or directly affect the output and quality of steel mill finished products. With the development of steelmaking enterprises and the improvement of process requirements, the current steelmaking production process control methods mainly have the following problems to be solved: control of molten steel carbon temperature in the converter production process, real-time monitoring and real-time prediction of furnace conditions, and determination of the best operating mode.

At present, most of the real-time prediction methods for molten steel quality in the converter steelmaking process only use a single detection method, fail to fully utilize the various detection methods of the converter, and fail to fully integrate and utilize the various common detection data of the converter. Multi-source data feature fusion collects, organizes and cleans data from different data sources to establish a consistent and complete data set, which can provide a more comprehensive data perspective. Multi-source data feature fusion can also reduce the errors and inaccuracies caused by a single data source. When a single technical method has abnormal or missing data, it is easy to cause real-time prediction inaccuracy. Multi-source data features can be used to identify and correct errors and duplications in the data, improve the prediction accuracy, improve the quality of molten steel, and reduce steelmaking costs.

Summary of the invention

In view of the problems existing in the prior art, the present invention provides a real-time prediction method for molten steel quality based on multi-source data feature fusion, aiming to make real-time prediction of molten steel quality in the converter steelmaking blowing process based on feature fusion of multi-source data, provide reference for on-site smelting production, improve production efficiency and product quality, and provide support for converter intelligence.

The technical solution of the present invention is:

A method for real-time prediction of molten steel quality based on multi-source data feature fusion, the method comprising the following steps:

Step 1: Build a training dataset;

Step 2: Use the training data set to train the existing XGBoost model, use the trained XGBoost model as the real-time prediction model for molten steel quality, and use the output of the model as the predicted value of molten steel temperature and molten steel carbon content at the next moment;

Step 3: The real-time blowing data of the new batch is transferred into the real-time prediction model of molten steel quality to predict the current real-time carbon content and temperature of the molten steel.

Further, according to the method for real-time prediction of molten steel quality based on multi-source data feature fusion, step 1 includes the following steps:

Step 1.1: Collect relevant production data, including the furnace entry condition data, oxygen blowing data, charging data, furnace mouth temperature data, audio data outside the fire wall, flue gas composition data of the converter flue, and carbon content and temperature of molten steel at the end of blowing measured by the auxiliary gun;

Step 1.2: Based on the relevant production data collected in step 1.1 and combined with the thermodynamic principle of converter reaction, the real-time carbon content and temperature of molten steel during blowing are calculated;

Step 1.3: Remove abnormal data and data corresponding to incomplete heats from the data obtained in steps 1.1 and 1.2, and the remaining data constitute the training data set.

Furthermore, according to the real-time prediction method of molten steel quality based on multi-source data feature fusion, the furnace entering condition data includes molten iron composition, molten iron temperature, molten iron weight, and scrap steel weight; the oxygen blowing data includes cumulative oxygen flow rate, cumulative oxygen blowing time, real-time oxygen flow rate, and real-time oxygen gun height; the feeding data includes cumulative limestone feeding amount, cumulative dolomite feeding amount, and cumulative cold material feeding amount; the flue gas composition data includes CO content and _CO2 content data.

Furthermore, according to the real-time prediction method of molten steel quality based on multi-source data feature fusion, the calculation formula of real-time molten steel carbon content is as follows:

w(C)=0.1×(∑C ₀ -∑C _de )/W _m (3)

Wherein, w(C) is the mass fraction of carbon in the molten pool, %; _∑C0 is the mass of molten iron C in the furnace charging condition data, kg; _∑Cde is the sum of the continuous decarburization amount, t.

Furthermore, according to the real-time prediction method of molten steel quality based on multi-source data feature fusion, the calculation formula of molten steel temperature is as follows:

Wherein, λ is the initial temperature correction coefficient, K; T is the molten steel temperature, K; _F10 is the _CO2 flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _F11 is the CO flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _f1 is the activity coefficient of C; _w1 is the C content in molten steel, %.

Furthermore, according to the real-time prediction method of molten steel quality based on multi-source data feature fusion, when the existing XGBoost model is trained using a training data set, the molten iron composition, molten iron temperature, molten iron weight, scrap steel weight, cumulative oxygen flow rate, cumulative oxygen blowing time, real-time oxygen flow rate, real-time oxygen gun height, cumulative limestone addition amount, cumulative dolomite addition amount, cumulative cold material addition amount, furnace mouth temperature data, audio data, flue gas CO content and flue gas _CO2 content measured by a flue gas analyzer in the training data set are used as training features; the real-time carbon content of the molten steel, the real-time temperature of the molten steel calculated in the training data set, and the carbon content and temperature of the molten steel at the end of blowing measured by the auxiliary gun are used as training labels.

Furthermore, according to the real-time prediction method of molten steel quality based on multi-source data feature fusion, the molten iron composition includes C content, P content, S content, Si content, and Mn content.

Compared with the prior art, the present invention has the following beneficial effects:

The method of the present invention utilizes multi-source data such as images, audio, flue gas data, raw material information, auxiliary material addition data, oxygen blowing parameters, etc. for feature fusion, uses thermodynamic mechanism, and performs model training based on XGBoost algorithm. Compared with the method of using single-source data for real-time prediction of molten steel quality, this method utilizes a variety of data commonly found in converters, has the advantages of strong anti-interference ability, good real-time performance, high precision, etc., and has guiding significance for actual production.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for real-time prediction of molten steel quality based on multi-source data feature fusion according to this embodiment;

FIG. 2 (a) is a comparison diagram of the predicted value and the actual value of the carbon content of the present embodiment; FIG. 2 (b) is a comparison diagram of the predicted value and the actual value of the molten steel temperature of the present embodiment;

FIG3 (a) is a comparison diagram of the molten steel temperature predicted by the smelting endpoint model and the actual value; FIG3 (b) is a comparison diagram of the carbon content value predicted by the smelting endpoint model and the actual value.

DETAILED DESCRIPTION

To facilitate understanding of the present application, a more comprehensive description of the present application will be given below with reference to the relevant drawings.

Figure 1 is a flow chart of a method for real-time prediction of molten steel quality based on multi-source data feature fusion according to this embodiment. As shown in Figure 1, the method for real-time prediction of molten steel quality based on multi-source data feature fusion comprises the following steps:

Step 1: Establish a training data set for the real-time prediction model of molten steel quality. First, collect relevant production data, including the furnace entry condition data, oxygen blowing data, charging data, furnace mouth temperature data, audio data outside the fire wall, flue gas composition data of the converter flue, and the carbon content of molten steel at the end of blowing (TSO-C) and molten steel temperature (TSO-T) measured by the auxiliary gun. Then, based on the thermodynamic principle of converter reaction, calculate the real-time molten steel carbon content and real-time molten steel temperature during the blowing process. Then, after data screening, form the training data set for the real-time prediction model of molten steel quality.

Step 1.1: Collect relevant production data;

The furnace entry condition data is specifically collected from the MES system in the steel plant, and the molten iron composition (C, P, S, Si, Mn), molten iron temperature, molten iron weight, and scrap steel weight in the collected data are used as the furnace entry condition data.

The oxygen blowing data is specifically transmitted through a TCP network to obtain real-time data on a PLC that controls the converter oxygen blowing equipment, wherein the accumulated oxygen flow, accumulated oxygen blowing time, real-time oxygen flow, and real-time oxygen gun height are used as oxygen blowing data.

The feeding data is specifically collected by the feeding system on the material type, preparation amount and addition amount of different silos, and the addition amount of the current furnace is accumulated, and the cumulative addition amount of limestone, dolomite and cold material is collected as the feeding data.

The temperature data is specifically obtained by pointing the thermal imager directly at the converter furnace mouth, acquiring the target temperature corresponding to each pixel at the furnace mouth, obtaining a two-dimensional temperature matrix of the furnace mouth flame, and extracting the maximum value in the two-dimensional temperature matrix as the furnace mouth temperature data of the current furnace.

The audio data is specifically collected by an audio acquisition device installed on the outside of the converter fire wall, and the sound intensity value is obtained as the audio data outside the fire wall.

The flue gas composition data is specifically obtained by measuring the converter flue gas composition through a flue gas analyzer on the converter flue, and extracting CO content and _CO2 content data as the flue gas composition data of the converter flue.

The carbon content and temperature of the molten steel at the blowing end point measured by the auxiliary gun are obtained by using the auxiliary gun after stopping the oxygen blowing.

Step 1.2: Combined with the thermodynamic principle of converter reaction, calculate the real-time carbon content and temperature of molten steel during blowing. The calculation process is as follows:

(1) Calculation of carbon content

According to the flue gas composition data collected in step 1.1, the decarbonization rate (kg/s) can be calculated according to the material balance:

Where, Q _gas is the flue gas flow rate, m ³ /s; x _CO is the mole fraction of carbon monoxide in the flue gas; is the molar fraction of carbon dioxide in the flue gas; _Wm is the mass of molten steel in the molten pool, t; w(C) is the mass fraction of carbon in the molten pool, %;

The total amount of continuous decarburization can be obtained by integrating the decarburization rate:

Where ∑C _de is the sum of the continuous decarburization amounts, t.

Combined with the C content of molten iron in the furnace feeding condition data collected in step 1.1, the real-time carbon content of molten steel can be calculated:

w(C)=0.1×(∑C ₀ -∑C _de )/W _m (3)

Wherein, _∑C0 is the mass of molten iron C in the furnace charging condition data, kg.

(2) Calculation of molten steel temperature

1) Calculation of molten steel basic temperature

During the converter smelting process, the molten pool reaction in the furnace will simultaneously generate two gases, CO and _CO2 , whose content ratio is related to the carbon content, oxygen content and temperature of the molten steel. The carbon content of the molten steel can be calculated in equation (3), so the oxygen content of the molten steel can be determined separately according to the heat balance equations of CO and _CO2 , and then the temperature of the molten steel can be calculated jointly.

Calculate the oxygen content of molten steel based on the thermal reaction balance equation of CO:

[C]+[O]＝CO

After finishing, we can get:

in, is the equilibrium reaction coefficient for the formation of CO; P1 is the equilibrium partial pressure of CO, Pa; P0 is the standard atmospheric pressure, Pa; _f1 , _f2 are the activity coefficients of C and O; _w1 , _w2 are the contents of C and O in molten steel, %; T is the molten steel temperature, K.

The calculation formula of carbon activity coefficient _f1 is as follows:

Calculate the oxygen content of molten steel based on the heat balance equation of CO ₂ :

[C] + 2 [O] = CO ₂

After finishing, we can get:

in, is the equilibrium reaction coefficient for the production of CO ₂ ; P2 is the equilibrium partial pressure of CO ₂ , Pa; P0 is the standard atmospheric pressure, Pa.

At the reaction interface, CO and _CO2 react simultaneously, so the oxygen content of molten steel calculated by their respective heat balance equations is equal, so equation (5) is equal to equation (8):

The basic molten steel temperature prediction formula can be obtained by sorting out:

Among them, _x1 and _x2 are the proportions of CO and _CO2 in the furnace gas at the molten pool reaction interface, respectively, %.

Because CO and _CO2 on the molten pool reaction interface will produce secondary combustion inside and outside the furnace before reaching the furnace gas analyzer, consuming CO to generate _CO2 , there is an error between the ratio of CO and _CO2 on the molten pool reaction interface and the ratio of CO and _CO2 detected by the furnace gas analyzer. Therefore, it is necessary to correct the ratio of CO and _CO2 detected by the furnace gas analyzer to satisfy the calculation of formula (10).

2) Calculation of CO ₂ flow rate generated by secondary combustion in the furnace

F ₃ = α(F ₁ -F ₂ ) ^β (11)

Among them, _F3 is the _CO2 flow rate generated by secondary combustion in the furnace, ^m3 /s; _F1 is the total flow rate of furnace gas, ^m3 /s; _F2 is the _N2 flow rate in the furnace gas, m3/ ^s ; α=0.2, β=0.75.

3) Calculation of _CO2 flow rate generated by secondary combustion outside the furnace

The law of conservation of mass is used to calculate the amount of air sucked into the flue at the furnace mouth, so as to know the _O2 flow rate sucked into the flue, and then the _CO2 flow rate of secondary combustion can be calculated.

The air flow rate at the furnace mouth is:

Among them, _F4 is the Ar flow rate in the flue, m ³ /s; _F5 is the air flow rate sucked into the furnace mouth, m ³ /s; _F6 is the Ar flow rate of the converter bottom blowing, m ³ /s; 0.934% is the Ar content in the air.

Therefore, the _O2 flow rate sucked into the furnace is:

_F7 ＝ _F5 ×20.95％ (13)

Among them, _F7 is the _O2 flow rate sucked into the furnace mouth, ^m3 /s; 20.95% is the _O2 content in the air.

The _CO2 flow rate produced by secondary combustion in the flue is:

_F9 = 2 × ( _F7 - _F8 ) (14)

Among them, _F8 is the _O2 flow rate in the flue gas, ^m3 /s; _F9 is the _CO2 flow rate produced by secondary combustion in the flue, ^m3 /s.

4) Corrected calculation of molten steel temperature

First, the ratio of CO and CO2 in the furnace gas inside and outside the furnace is corrected:

_F10 ＝ _F12 - _F3 - _F9 (15)

_F11 ＝ _F13 + _F3 + _F9 (16)

Among them, _F10 is the _CO2 flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _F12 is the _CO2 flow rate in the flue, ^m3 /s; _F3 is the _CO2 flow rate generated by the secondary combustion in the furnace, ^m3 /s; _F9 is the _CO2 flow rate generated by the secondary combustion in the flue, ^m3 /s; _F11 is the CO flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _F13 is the CO flow rate in the flue, ^m3 /s.

Therefore, the CO ratio x ₁ ' and CO ₂ ratio x ₂ ' at the molten pool reaction are:

Then the temperature calculation formula after correction of formula (10) is:

Wherein, λ is the initial temperature correction coefficient, K; the specific value of λ is given based on experience.

Step 1.3: Eliminate abnormal and incomplete heats. Delete the heats with missing data and abnormal smelting time, and screen the heats according to the molten iron composition range and auxiliary gun measurement value range set by manual experience. The processed data is used as the training data set for the real-time prediction model of molten steel quality; Table 1 shows the parameter values of the training data set of the real-time prediction model of molten steel quality.

Table 1 Parameter values of the training data set for the real-time prediction model of molten steel quality

Step 2: Take the molten iron C content, molten iron P content, molten iron S content, molten iron Si content, molten iron Mn content, molten iron temperature, molten iron weight, scrap steel weight, cumulative oxygen flow, cumulative oxygen blowing time, real-time oxygen flow, real-time oxygen lance height, cumulative limestone feeding amount, cumulative dolomite feeding amount, cumulative cold material feeding amount, furnace mouth temperature data, audio data, flue gas CO content and flue gas _CO2 content measured by the flue gas analyzer obtained in step 1 as training features, and take the calculated real-time C content of molten steel, real-time temperature of molten steel, and carbon content (TSO-C) and temperature (TSO-T) of molten steel at the end of blowing measured by the auxiliary lance as training labels, use the training data set constructed in step 1 to train the XGBoost model, and obtain a real-time prediction model for molten steel quality, and use the output of the model as the predicted value of molten steel temperature and carbon content at the next moment.

XGBoost is a machine learning algorithm based on the Gradient Boosting Decision Tree, which is used to solve classification and regression problems. It uses the cumulative method to predict the real-time temperature and carbon content of molten steel. The construction process of this model includes:

(1) Establishing the objective function

The objective function of XGBoost is:

where _yi is the actual value of the ith sample, is the predicted value of the ith sample, It is the sum of the loss functions of all samples. The complexity of all t trees is summed up and added to the objective function as a regularization term to prevent overfitting of the t models.

The prediction value of the XGBoost model, i.e., the real-time temperature of the molten steel or the real-time carbon content of the molten steel, is calculated by the cumulative method, so the prediction value of the first t trees for the i-th sample is for:

in is the prediction value of the first t-1 prediction trees for the i-th sample, and f _t ( _xi ) is the prediction value of the t-th tree for the i-th sample. Therefore, the objective function of the model can be rewritten as:

Among them, _gi is the first-order derivative of the loss function, and _hi is the second-order derivative of the loss function.

(2) Building a regression tree

Use a greedy algorithm to select the feature with the greatest benefit (the regression problem is the sum of the mean absolute errors of the sum of the partition sets) as the split feature, use the best splitting point of the feature as the splitting position, split two new leaf nodes on the left and right at the node, and associate the corresponding sample set with each new node.

The details are as follows: Starting from the depth of the tree is 0:

1) For each leaf node, enumerate all available features, namely, molten iron C content, molten iron P content, molten iron S content, molten iron Si content, molten iron Mn content, molten iron temperature, molten iron weight, scrap weight, cumulative oxygen flow, cumulative oxygen blowing time, real-time oxygen flow, real-time oxygen gun height, cumulative limestone feeding amount, cumulative dolomite feeding amount, cumulative cold material feeding amount, furnace mouth temperature data, audio data, flue gas CO content measured by flue gas analyzer, flue gas _CO2 content;

2) For each feature, the training samples belonging to the node are sorted in ascending order according to the feature value, the best split point of the feature is determined by linear scanning, and the split benefit of the feature is recorded;

3) Select the feature with the largest benefit as the split feature, use the best split point of the feature as the split position, split two new leaf nodes on the left and right sides of the node, and associate the corresponding sample set for each new node (each new node is associated with a subset of the total training data set, and this subset is divided according to the split condition);

4) Go back to step 1) and enumerate all available features for each leaf node, recursively executing until a specific condition is met;

Step 3: The real-time blowing data of the new batch is transferred into the real-time prediction model of molten steel quality trained in step 2 to predict the current real-time carbon content and temperature of the molten steel;

1) Furnace feeding condition data: molten iron composition (C, P, S, Si, Mn), molten iron temperature, molten iron weight, scrap steel weight.

2) Oxygen blowing data: cumulative oxygen flow, cumulative oxygen blowing time, real-time oxygen flow, and real-time oxygen gun height. 3) Feeding data: cumulative feeding amount of lime, dolomite, and cold material. 4) Furnace mouth temperature data: the furnace mouth image captured by the thermal imager is intercepted and features are extracted to obtain the maximum value of the temperature matrix. 5) Audio data outside the fire wall: sound intensity value collected by the audio acquisition device 6) Flue gas composition data of the converter flue: flue gas composition data obtained by the flue gas analyzer installed on the flue. The above real-time blowing data is passed into the model trained in step 2 to calculate the current real-time carbon content and temperature of the molten steel.

The carbon temperature prediction curve calculated based on the actual production data of a steel plant is shown in Figure 2(a) and Figure 2(b). After verification by actual production data, the comparison results of the predicted molten steel temperature and the actual value obtained by the endpoint model of multiple furnaces are shown in Figure 3(a), and the comparison results of the endpoint carbon content value and the actual value are shown in Figure 3(b). The average error of the endpoint temperature prediction is 11.25°C, and the average error of the endpoint carbon content prediction is 0.034%. It has been verified that the method of the present invention can provide a reference for on-site operators, which is helpful to improve the carbon temperature hit rate at the end of converter molten steel smelting and improve production efficiency.

It should be understood that, inspired by the technical concept of the present invention, those skilled in the art may make various improvements or changes based on the above content without departing from the content of the present invention, which still fall within the protection scope of the present invention.

Claims (7)

1. A method for real-time prediction of molten steel quality based on multi-source data feature fusion, characterized in that the method comprises the following steps: Step 1: Build a training dataset; Step 2: Use the training data set to train the existing XGBoost model, use the trained XGBoost model as the real-time prediction model for molten steel quality, and use the output of the model as the predicted value of molten steel temperature and molten steel carbon content at the next moment; Step 3: The real-time blowing data of the new batch is transferred into the real-time prediction model of molten steel quality to predict the current real-time carbon content and temperature of the molten steel. 2. The method for real-time prediction of molten steel quality based on multi-source data feature fusion according to claim 1, characterized in that step 1 comprises the following steps: Step 1.1: Collect relevant production data, including the furnace entry condition data, oxygen blowing data, charging data, furnace mouth temperature data, audio data outside the fire wall, flue gas composition data of the converter flue, and carbon content and temperature of molten steel at the end of blowing measured by the auxiliary gun; Step 1.2: Based on the relevant production data collected in step 1.1 and combined with the thermodynamic principle of converter reaction, the real-time carbon content and temperature of molten steel during blowing are calculated; Step 1.3: Remove abnormal data and data corresponding to incomplete heats from the data obtained in steps 1.1 and 1.2, and the remaining data constitute the training data set. 3. According to the real-time prediction method of molten steel quality based on multi-source data feature fusion according to claim 1, it is characterized in that the furnace entry condition data includes molten iron composition, molten iron temperature, molten iron weight, and scrap steel weight; the oxygen blowing data includes cumulative oxygen flow rate, cumulative oxygen blowing time, real-time oxygen flow rate and real-time oxygen gun height; the feeding data includes cumulative limestone feeding amount, cumulative dolomite feeding amount and cumulative cold material feeding amount; the flue gas composition data includes CO content and _CO2 content data. 4. The method for real-time prediction of molten steel quality based on multi-source data feature fusion according to claim 1, characterized in that the calculation formula of real-time molten steel carbon content is as follows: w(C)=0.1×(∑C ₀ -∑C _de )/W _m (3) Wherein, w(C) is the mass fraction of carbon in the molten pool, %; _∑C0 is the mass of molten iron C in the furnace charging condition data, kg; _∑Cde is the sum of the continuous decarburization amount, t. 5. The method for real-time prediction of molten steel quality based on multi-source data feature fusion according to claim 1, characterized in that the calculation formula of molten steel temperature is as follows: Wherein, λ is the initial temperature correction coefficient, K; T is the molten steel temperature, K; _F10 is the _CO2 flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _F11 is the CO flow rate directly generated by the molten pool decarburization reaction, ^m3 /s; _f1 is the activity coefficient of C; _w1 is the C content in molten steel, %. 6. The method for real-time prediction of molten steel quality based on multi-source data feature fusion according to claim 2 is characterized in that when the existing XGBoost model is trained using a training data set, the molten iron composition, molten iron temperature, molten iron weight, scrap steel weight, cumulative oxygen flow rate, cumulative oxygen blowing time, real-time oxygen flow rate, real-time oxygen gun height, cumulative limestone feeding amount, cumulative dolomite feeding amount, cumulative cold material feeding amount, furnace mouth temperature data, audio data, flue gas CO content and flue gas _CO2 content measured by a flue gas analyzer in the training data set are used as training features; the real-time carbon content of the molten steel, the real-time temperature of the molten steel calculated in the training data set, and the carbon content and temperature of the molten steel at the blowing end point measured by the auxiliary gun are used as training labels. 7. The real-time prediction method for molten steel quality based on multi-source data feature fusion according to claim 3 or 6 is characterized in that the molten iron composition includes C content, P content, S content, Si content, and Mn content.