Spectral feature information extraction method for converter steelmaking endpoint prediction

Abstract

The present invention discloses a method for extracting spectral characteristic information for predicting the endpoint of converter steelmaking, and belongs to the field of photoelectric detection technology. By collecting spectral images and analyzing characteristic parameters, the potential laws of the blowing process can be found, eliminating the need to analyze the physical and chemical reaction mechanisms of the steelmaking process. First, polynomial fitting is used to find the background information of the spectrum; second, after subtracting the background, the emission peak of the characteristic atom is found through Gaussian fitting; finally, the spectral difference spectrum in the time domain of the blowing endpoint period is peak-searched to find the three most obvious small characteristic values. The extracted characteristic parameters use the SVR regression model to predict the carbon content. Experiments show that the prediction accuracy of the steelmaking endpoint can reach more than 90%.

Description

Technical Field

[0001] The invention belongs to the field of photoelectric detection, and in particular is a method for extracting spectral characteristic information for predicting the end point of converter steelmaking. Background Art

[0002] Converter steelmaking is widely used in the steelmaking industry. Converter endpoint control is a key factor in determining steelmaking quality. Due to the instability of added raw materials, complex chemical reactions during blowing and the strict range of steel grades required for blowing, accurate control of the carbon content at the converter endpoint is a difficult problem that needs to be solved urgently at home and abroad. According to the carbon content, carbon steel can be divided into high carbon steel (carbon content 0.6-2%), medium carbon steel (carbon content 0.3-0.6%), and low carbon steel (carbon content 0.03-0.3%). The small difference in carbon content among the three types brings great difficulty to control. Based on this, many studies have been carried out around the prediction of carbon content and temperature values ​​at the converter steelmaking endpoint.

[0003] The control of the converter end point has mainly gone through several development stages, including manual experience control, static control and dynamic control.

[0004] Manual experience control mainly relies on the workers' experience to judge the end point, observe the spark changes, flame shape and strength, steel samples, etc. This method is widely used in steelmaking control at home and abroad, but this method is very inaccurate and requires a lot of manpower and material resources, and there are major problems.

[0005] Static control selects the endpoint carbon of molten steel as the target value, and calculates the amount of various raw materials and oxygen added based on known raw materials and thermal balance principles. According to different modeling methods, static control is divided into three types: empirical model, statistical model, and mechanism model. The static control method improves the hit rate of the endpoint, but because the steelmaking process is a complex reaction process, it is difficult to calculate accurately and cannot be dynamically corrected, and has certain limitations.

[0006] Currently, the main dynamic control methods include the auxiliary gun method and the furnace gas method: the auxiliary gun method inserts an auxiliary gun into the molten steel at the end of blowing to obtain the temperature and carbon content of the molten steel at this time, and corrects the calculation results of the static model based on the data, thereby improving the degree of automation. However, the auxiliary gun control system has high requirements on the converter type and investment cost, and is difficult to promote to small and medium-sized converter plants; the furnace gas method uses a mass spectrometer to detect furnace gas components such as carbon monoxide and carbon dioxide overflowing from the furnace mouth, and corrects parameters in real time through a dynamic model to improve control accuracy. However, the instruments used in this method are expensive and difficult to maintain, and are also difficult to promote to small and medium-sized converter plants.

[0007] Since the 1980s, the analysis and research of furnace mouth radiation information has been gradually carried out. Through the analysis of spectral information, flame images, etc., a non-contact method has been constructed to predict the carbon content of the converter end point, and the modeling direction has also begun to turn to the intelligent method modeling of neural networks, avoiding the unclear reaction mechanism in the furnace and using a black box model to improve the end point hit rate. Spectral information is easy to obtain using a spectrometer and contains a lot of information, while continuous shooting of flame images is difficult, the amount of data is large, and the algorithm is complex. Therefore, the prediction model for end point control based on spectral images has a good application in end point control.

[0008] In summary, the non-contact prediction method based on spectral image has a higher hit rate and can significantly reduce costs compared with traditional empirical methods and static control methods. However, there is still room for research on spectral feature information extraction and neural network modeling methods. The present invention proposes a new spectral feature information extraction method for predicting the end point of converter steelmaking. Summary of the invention

[0009] The purpose of the present invention is to provide a spectral feature information extraction method for predicting the endpoint of converter steelmaking, which can extract the furnace mouth spectral features with strong distinguishing properties under different molten steel carbon contents and determine the position of the steelmaking endpoint more accurately and quickly.

[0010] The technical solution to achieve the purpose of the present invention is: a method for extracting spectral feature information for predicting the endpoint of converter steelmaking, the method comprising the following steps:

[0011] Step 1: Collect the spectral image of the converter steelmaking end point furnace mouth to construct a training data set, and the image label corresponds to the carbon content value of the molten steel detected by the auxiliary gun at the end point;

[0012] Step 2, select N frames of spectral images before the end time for feature extraction, including fitting the spectrum lines using low-order polynomials, Gaussian fitting of the spectrum using the least squares method, and peak analysis of the differential spectrum.

[0013] Furthermore, the step 2 of selecting N frames of spectral images before the end time for feature extraction specifically includes:

[0014] Step 2-1, smoothing the spectral image in the time domain with a step size of N / 4 frames, and taking the smoothed image of the last N / 4 frames as the end spectral image;

[0015] Step 2-2, performing n-term fitting on the endpoint spectrum image to obtain background information, namely, gray body radiation spectrum, and the fitting mathematical expression is:

[0016]

[0017] The coefficients of the spectra fitted from different batches are different, and the peak value and effective width of the polynomial are obtained as the characteristic values ​​x1 and x2 respectively;

[0018] In the formula, a k are the coefficients in the polynomial function, x k is the independent variable, p n (x) is the fitting function value;

[0019] Step 2-3, Gaussian fitting is performed on the spectrum after subtracting the background information from the endpoint spectrum image, and the emission peaks of the two characteristic atoms are obtained by using the least square method. The mathematical expression is:

[0020] G i (x) = A i *exp((xB i )^2 / C i ^2)

[0021] The obtained emission peak values ​​are taken logarithmically as eigenvalues ​​x3 and x4 respectively;

[0022] In the formula, A i , B i With C i are real constants, and A i >0;

[0023] Step 2-4, performing time domain difference processing on the four smoothed images in step 2-1 to obtain a difference spectrum;

[0024] Step 2-5, performing peak fitting on the differential spectrum to obtain the wavelength ranges of three groups of small characteristic peaks: characteristic peak 1-[667.88, 671.63], characteristic peak 2-[720.59, 732.83], characteristic peak 3-[817.39, 818.57], in nm;

[0025] Step 2-6, respectively obtain the integral values ​​of the discrete light intensity values ​​on the endpoint spectrum corresponding to the above three wavelength ranges as characteristic values ​​x5, x6, and x7.

[0026] Compared with the prior art, the present invention has the following significant advantages: the present invention proposes a method for extracting spectral feature information for predicting the endpoint of converter steelmaking by analyzing spectral feature information. The present invention adopts three schemes to extract different feature parameters. Among them, differential processing of the spectrum can more conveniently and accurately find the position of small feature peaks compared with the prior art, and taking the integrated area as the feature value can more obviously reflect the dynamic changes of the peak. Experiments show that the method of the present invention can extract spectral features with strong discrimination under different carbon contents, and the extracted features have achieved good results in endpoint judgment, thereby proving its effectiveness.

[0027] The present invention is further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 The present invention is a flow chart of the spectral feature information extraction method for converter steelmaking endpoint prediction.

[0029] Figure 2 Schematic diagram of characteristic peaks of furnace mouth spectrum image. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0031] In the process of converter steelmaking, the carbon content of molten steel at the end point is closely related to the spectral information at the furnace mouth, so the effective extraction of spectral feature information is the key to accurate prediction.

[0032] In one embodiment, in combination Figure 1 , provides a method for extracting spectral feature information for predicting the endpoint of converter steelmaking, the method comprising the following steps:

[0033] Step 1: Collect the spectral image of the converter steelmaking end point furnace mouth to construct a training data set, and the image label corresponds to the carbon content value of the molten steel detected by the auxiliary gun at the end point;

[0034] Here, when the input data is high-dimensional data, the amount of calculation will increase geometrically, resulting in slow operation speed. Therefore, the spectral data must be reduced in dimension to reduce the amount of calculation. The spectrum obtained in the experiment contains the radiation spectrum of molten steel ash and the emission spectrum of carbon atoms. Based on the law of spectral radiation, the present invention extracts stable features and unstable features from the changes in the spectral image in the late stage of blowing as the input of the neural network model.

[0035] Step 2, select N frames of spectral images before the end time for feature extraction, including fitting the spectrum line using low-order polynomials, Gaussian fitting of the spectrum using the least squares method, and peak analysis of the differential spectrum. Here, the spectrum line is fitted using low-order polynomials to characterize the spectral information to the greatest extent and prevent overfitting. The continuous spectrum in the spectral image is an important feature to describe the overall change and can reflect the change in the flame intensity at the furnace mouth.

[0036] Furthermore, in one of the embodiments, the endpoint time in step 1 is the time when the carbon content value of the molten steel is collected.

[0037] Furthermore, in one of the embodiments, the spectral image is collected by installing a telescopic collimation system in front of the grating spectrometer.

[0038] Further, in one embodiment, the step 2 of selecting N frames of spectral images before the end time for feature extraction specifically includes:

[0039] Step 2-1, smoothing the spectral image in the time domain with a step length of N / 4 frames, the curve trend is obviously highlighted, and the smoothed image of the last N / 4 frames is used as the end spectral image;

[0040] Step 2-2, performing n-term fitting on the endpoint spectrum image to obtain background information, namely, gray body radiation spectrum, and the fitting mathematical expression is:

[0041]

[0042] The coefficients of the spectra fitted from different batches are different, and the peak value and effective width of the polynomial are obtained as the characteristic values ​​x1 and x2 respectively;

[0043] In the formula, a k are the coefficients in the polynomial function, x k is the independent variable, p n (x) is the fitting function value;

[0044] Here, in the actual fitting process, it is found that the emission peak may be very wide, and the peak height of the continuous spectrum is easily affected by the peak value of the emission peak and increased. Intuitively, the spectrum obtained by removing the emission peak band from the original spectrum can be fitted, but the fitting curve is easily affected by the protruding part of the right-hand base radiation, which reduces the skewness and makes the fitting of the subsequent emission peak inaccurate. Therefore, a relatively stable area is taken between the two emission peaks. The area is fitted with the least squares method to find the mean, and the low-order polynomial curve fitted by the entire spectrum is modulated with this mean. Because the selected area is only a stable area that is not an emission peak, rather than the peak of the true continuous spectrum, it is also necessary to multiply the peak value of the spectrum obtained by fitting multiple furnaces of the original spectrum to remove the emission peak band by dividing the peak value of the spectrum obtained by dividing the mean value corresponding to the middle section of the two emission peaks of the furnace spectrum by a constant as a priori, which is about 1.03. The modulation curve is selected as an estimate of the continuous spectrum, and its peak value and effective width are used as input parameters x1 and x2.

[0045] Step 2-3, Gaussian fitting is performed on the spectrum after subtracting the background information from the endpoint spectrum image, and the emission peaks of the two characteristic atoms are obtained by using the least square method. The mathematical expression is:

[0046] G i (x) = A i *exp((xB i )^2 / Ci ^2)

[0047] The obtained emission peak values ​​are taken logarithmically as eigenvalues ​​x3 and x4 respectively;

[0048] In the formula, A i , B i With C i are real constants, and A i >0;

[0049] Here, according to the steel decarburization theory, the carbon content in the furnace is significantly reduced in the later stage of blowing, and the decarburization rate is proportional to the carbon content inside the molten pool.

[0050] Subtract the original spectrum from the modulation curve, use SG with a short step size to smooth out small burrs to prevent Gaussian fitting errors, and then use the least squares method to perform Gaussian fitting on the spectrum to obtain the wavelengths corresponding to the emission peaks, which are about 589.8nm and 767.6nm. Figure 2 As shown;

[0051] Step 2-4, perform time domain difference processing on the four smoothed images in step 2-1 (the peak characteristic change of the spectrum before the end point can be obtained, the overlapping part of the spectrum has obvious signs of aggravation, and the background part is eliminated, so that the position and amount of spectrum change can be determined) to obtain a differential spectrum;

[0052] Step 2-5, performing peak fitting on the differential spectrum to obtain the wavelength ranges of three groups of small characteristic peaks: characteristic peak 1-[667.88, 671.63], characteristic peak 2-[720.59, 732.83], characteristic peak 3-[817.39, 818.57], in nm;

[0053] Step 2-6, respectively obtain the integral values ​​of the discrete light intensity values ​​on the endpoint spectrum corresponding to the above three wavelength ranges as characteristic values ​​x5, x6, and x7.

[0054] Further preferably, in one of the embodiments, N is 20 and n is 6.

[0055] A carbon content prediction method based on the above-mentioned spectral feature information extraction method for converter steelmaking endpoint prediction, the method comprising:

[0056] Step 1, obtaining spectral characteristic data of different batches to form a sample data set;

[0057] Step 2, dividing the sample data set into a training set (90%) and a test set (10%);

[0058] Step 3, using the training set to train a support vector machine regression model to obtain a carbon content prediction model;

[0059] Step 4: for the furnace whose carbon content is to be predicted, collect the spectral image of its furnace mouth, extract the spectral feature data, and input it into the carbon content prediction model to obtain the carbon content.

[0060] Furthermore, in one of the embodiments, step 1 also includes normalizing the data in the sample data set.

[0061] Here, since the numerical variation range of each feature is different, the contribution of features with smaller values ​​may not be obvious, affecting the overall prediction effect, so normalization is required. There are usually three normalization processing methods: linear transformation normalization, regularization normalization, and standard normalization. Taking into account the training effect and data operability, the present invention chooses to use the regularization normalization method to process the input data. The data after regularization is between 0-1.

[0062] Furthermore, in one of the embodiments, in the process of training the support vector machine regression model using the training set in step 3, a sparrow search algorithm is used to find the optimal kernel function.

[0063] The present invention mainly starts with the analysis of the characteristics of the furnace mouth spectral image, extracts effective characteristic information from the spectral images with highly similar carbon contents, and establishes a relatively accurate correspondence between the spectral image and the carbon content of molten steel. The present invention constructs a furnace mouth spectral feature extraction model that combines stable features with unstable features. Experiments show that the method of the present invention can extract spectral features with strong distinguishability under different carbon contents, and the extracted features have achieved good results in endpoint judgment, thereby proving its effectiveness.

[0064] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for extracting spectral feature information for predicting the endpoint of converter steelmaking, characterized in that: The method comprises the following steps: Step 1: Collect the spectral image of the converter steelmaking end point furnace mouth to construct a training data set, and the image label corresponds to the carbon content value of the molten steel detected by the auxiliary gun at the end point; Step 2, selecting N frames of spectral images before the endpoint moment for feature extraction, including fitting the spectrum line using a low-order polynomial, Gaussian fitting the spectrum using the least squares method, and peak analysis of the differential spectrum; the selecting N frames of spectral images before the endpoint moment for feature extraction specifically includes: Step 2-1, smoothing the spectral image in the time domain with a step size of N / 4 frames, and taking the smoothed image of the last N / 4 frames as the end spectral image; Step 2-2, performing n-term fitting on the endpoint spectrum image to obtain background information, namely, gray body radiation spectrum, and the fitting mathematical expression is: The coefficients of the spectra fitted from different batches are different, and the peak value and effective width of the polynomial are obtained as the characteristic values ​​x1 and x2 respectively; In the formula, a k are the coefficients in the polynomial function, x k is the independent variable, p n (x) is the fitting function value; Step 2-3, Gaussian fitting is performed on the spectrum after subtracting the background information from the end point spectrum image, and the emission peaks G of the two characteristic atoms are obtained by using the least square method. i (x), the mathematical expression is: G i (x)=A i *exp((x-B i )^2 / C i ^2) The obtained emission peak values ​​are taken logarithmically as eigenvalues ​​x3 and x4 respectively; In the formula, A i , B i With C i are real constants, and A i >0; Step 2-4, performing time domain difference processing on the four smoothed images in step 2-1 to obtain a difference spectrum; Step 2-5, performing peak fitting on the differential spectrum to obtain the wavelength ranges of three groups of small characteristic peaks: characteristic peak 1-[667.88, 671.63], characteristic peak 2-[720.59, 732.83], characteristic peak 3-[817.39, 818.57], in nm; Step 2-6, respectively obtain the integral values ​​of the discrete light intensity values ​​on the endpoint spectrum corresponding to the above three groups of wavelength ranges as characteristic values ​​x5, x6, and x7.

2. The spectral feature information extraction method for converter steelmaking endpoint prediction according to claim 1, characterized in that: The endpoint time in step 1 is the time when the carbon content value of the molten steel is collected.

3. The method for extracting spectral characteristic information for predicting the endpoint of converter steelmaking according to claim 2, characterized in that: The spectral image is collected by installing a telescopic collimation system in front of the grating spectrometer.

4. The method for extracting spectral characteristic information for predicting the endpoint of converter steelmaking according to claim 1, characterized in that: The N is 20 and n is 6.

5. A carbon content prediction method based on the method according to any one of claims 1 to 4, characterized in that: The method comprises: Step 1, obtaining spectral characteristic data of different batches to form a sample data set; Step 2, dividing the sample data set into a training set and a test set; Step 3, using the training set to train a support vector machine regression model to obtain a carbon content prediction model; Step 4: for the furnace whose carbon content is to be predicted, collect the spectral image of its furnace mouth, extract the spectral feature data, and input it into the carbon content prediction model to obtain the carbon content.

6. The carbon content prediction method according to claim 5, characterized in that: Step 1 also includes normalizing the data in the sample data set.

7. The carbon content prediction method according to claim 6, characterized in that: The normalization process specifically adopts a regular normalization method.

8. The carbon content prediction method according to claim 7, characterized in that: In the process of training the support vector machine regression model using the training set described in step 3, the sparrow search algorithm is used to find the optimal kernel function.

Images