CN106153551A - Converter steel-smelting molten steel carbon content based on SVM online Real-time and Dynamic Detection system - Google Patents

Abstract

The invention provides an SVM-based online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking, comprising: a telescopic optical system configured to collect flame image information at the mouth of a steelmaking furnace in real time; a spectrometer configured to receive information from Described flame image information, and obtain the flame spectral information of flame image information; Based on the SVM terminal control device, have a computing unit and the central control unit that controls computing unit operation, computing unit is used for according to the flame spectral information that obtains in real time through SVM The carbon content dynamic prediction model detects the carbon content in real time; the telescopic optical system includes an objective lens with a common optical axis, an eyepiece, and a field of view diaphragm. The field of view diaphragm is arranged in the optical imaging path formed by the objective lens and the eyepiece to adjust Field of view for flame detection at the furnace mouth. The detection system proposed by the invention has high field test accuracy, is not affected by external environmental factors, and has strong anti-interference ability.

Description

On-line real-time dynamic detection system for carbon content in molten steel in converter steelmaking based on SVM

technical field

Various aspects of the present invention relate to the technical field of converter steelmaking, especially the real-time monitoring of carbon content in molten steel during converter steelmaking, and specifically relate to an online real-time dynamic detection system for carbon content in molten steel in converter steelmaking based on SVM.

Background technique

The mainstream steelmaking technology in the world today is converter steelmaking, whose output accounts for more than 70% of the total steel output. The most important link in the converter steelmaking process is the final end point control, which is directly related to the quality of the final molten steel. Since the emergence of the converter steelmaking method, the end point control of the converter steelmaking has mainly experienced four development stages: manual experience control, static model control, dynamic model control and optical information control.

Manual empirical control, that is, empirical steelmaking, uses the means of thermocouple temperature measurement and carbon determination and rapid analysis of furnace front sampling to carry out manual empirical judgment control on the end point of the converter under normal blowing conditions. The carbon-oxygen reaction rate is an important basis for dividing the three stages, and the intensity of the carbon-oxygen reaction and the temperature of molten steel can be reflected by the flame at the furnace mouth. Steelmaking operators comprehensively judge the end point of steelmaking by observing the furnace mouth flame, sparks and oxygen supply time. However, only relying on the naked eye observation of the operator has problems such as a low hit rate at the end point and high labor intensity for the workers.

Static model control is to conduct statistical analysis on the initial data of previous converter blowing according to the principle of statistics, calculate the initial conditions required for blowing, and use these conditions to carry out the blowing process. Generally speaking, compared with manual empirical control, static model control can more effectively use the initial conditions of blowing process for quantitative calculation and control. Static model control can find the best raw material ratio according to the raw material conditions, and determine the smelting plan according to the actual ingredients, so as to overcome the randomness and inconsistency of empirical control. The existing static models include three types: mechanism model, statistical model and incremental model. In practical applications, these three models are often combined to improve the accuracy of end-point control. However, since the static model control does not consider the dynamic information in the blowing process, online tracking and real-time correction cannot be performed, so the accuracy is greatly limited.

The dynamic model control is mainly the dynamic control method of the sub-lance. On the basis of the static model, the sub-lance is used to detect the molten steel in the converter. According to the detection results, the initial parameters are corrected to obtain an accurate end point. Especially in recent years, with the application of artificial neural network research on dynamic model control methods, the problem of traditional static model control ignoring the dynamic information in the blowing process has been overcome, the accuracy of prediction has been further improved, and the end point prediction results The hit rate has been further improved, and the automation of steelmaking has been greatly improved. However, its cost is relatively high, and the converter needs to be modified, so it is not suitable for general small and medium-sized converters.

The traditional method is either inaccurate in judging the end point, or the cost is high and the adaptability is limited. Therefore, with the development of steelmaking technology and the progress of related technologies, people continue to try to apply more effective and accurate methods in the end point control technology. In the 1980s, a new endpoint control method for judging the endpoint of converter steelmaking using the optical information of the converter mouth appeared. For example, an optical detector that uses the changes that occur when the infrared laser penetrates the furnace gas to measure the composition of the furnace gas to judge the end point. The detector judges the end point by detecting the changes that occur when the laser passes through the furnace gas. The content of carbon monoxide in the furnace gas is controlled at the end point according to the change in the composition of carbon monoxide in the furnace gas. In experience or dynamic model control, what cannot be ignored is that the operator needs to obtain different degrees of information from the change of the flame, which is actually the flame's aperture, spectral distribution and image information of the flame. With the continuous development of photoelectric devices and the continuous maturity of optical processing methods, optical detection technology has been greatly developed, and optical control methods have also been applied to the end point control of converter steelmaking. Such as the molten steel radiation spectral information detection method proposed by Zhang Jinjin, Shi Yanjie, etc., the furnace mouth flame light intensity information detection method proposed by the Bethlehem Steel Company of the United States, and the flame image information detection method proposed by Wei Chengye, Yan Jianhua, etc.

Although the research on the steelmaking endpoint control theory continues to deepen, the cost of these methods is extremely high, the cost of detection and analysis equipment is extremely expensive, and the installation and maintenance are very inconvenient, so they are only used in some powerful steel enterprises. . In most small and medium iron and steel enterprises, a single empirical control or static model control is still the main method. Although the latest optical information control method provides some valuable ideas and application directions, due to the limitation of production scale and production conditions, especially the complex and harsh steelmaking production environment, in terms of optical information collection, the anti-interference ability Weak, unable to quickly and continuously extract the required parameter information, so it is difficult for some small and medium iron and steel enterprises to accept.

Therefore, there is an urgent need to develop an accurate online real-time steelmaking endpoint control scheme suitable for small and medium-sized iron and steel enterprises and small and medium-sized converters.

Contents of the invention

The purpose of the present invention is to provide an online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking based on SVM. The method is designed based on the support vector machine algorithm and has the advantages of high carbon content prediction accuracy, non-contact, strong anti-interference ability, and easy operation. , thus solving the problem of online dynamic detection of carbon content in converter steelmaking.

The above objects of the invention are achieved by the technical features of the independent claims, which the dependent claims develop in an alternative or advantageous manner.

In order to achieve the above object, the present invention proposes a SVM-based online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking. The detection system includes:

The telescopic optical system is configured to collect the flame image information of the steelmaking furnace mouth in real time;

The spectrometer is configured to receive flame image information from the telescopic optical system through an optical fiber, and obtain flame spectral information of the flame image information;

An end point control device based on SVM, the device has a computing unit and a central control unit that controls the operation of the computing unit, and the computing unit is configured to carry out steelmaking through the SVM carbon content dynamic prediction model according to the flame spectrum information obtained in real time Real-time detection of carbon content in water;

It is characterized by:

The telescopic optical system includes an objective lens with a common optical axis, an eyepiece, and a field of view diaphragm independent of the objective lens and the eyepiece. The field of view diaphragm is arranged in the optical imaging path formed by the objective lens and the eyepiece for adjusting the furnace Field of view for flame detection.

In a further embodiment, the field diaphragm is located on the focal plane of the objective lens, or, the field diaphragm is located behind the eyepiece and close to the optical fiber.

In a further embodiment, the field of view diaphragm is a variable field of view diaphragm.

In a further embodiment, the objective lens is a double separation lens, which is formed by the distribution of a positive lens and a negative lens with a common optical axis; the eyepiece is a Kenel eyepiece, which consists of a single lens and a doublet lens. distribution of the axes.

In a further embodiment, the SVM-based endpoint control device, wherein the computing unit is realized by one of FPGA and CPLD, the SVM carbon content dynamic prediction model is burned in the FPGA or CPLD, and upon receiving Automatically detect the carbon content after receiving the flame spectrum information.

In a further embodiment, the SVM carbon content dynamic prediction model includes:

A parameter building block for constructing a characteristic parameter characterizing the change of carbon content in the furnace according to the input flame spectrum information;

A carbon content dynamic prediction module for performing carbon content prediction based on the constructed characteristic parameters; and

An output module for predicting result output;

Wherein, the parameter construction module is configured to construct characteristic parameters in the following manner:

The shape of the spectrum at a wavelength of 600nm is a raised peak, and the characteristic parameter a ₁ is the normalized value of the light intensity here;

The peak of the spectral shape raised at 770nm is a _double peak, and the characteristic parameter a2 is the normalized mean value of the light intensity at the wavelength of 770nm and 772nm;

The continuous spectrum in the middle of the two sharp peaks changes drastically, and the spectral line is divided into three segments on average, and the light intensity of each segment is normalized and averaged to obtain three characteristic parameters a ₃ , a ₄ , and a ₅ ;

The ratio of the peak wavelength λ of the spectrum in the spectral distribution to the maximum value Tmax of the detection range of the spectrometer is used as the sixth parameter: a ₆ .

In a further embodiment, in the SVM carbon content dynamic prediction model, the carbon content dynamic prediction module for performing carbon content prediction based on the constructed characteristic parameters implements the establishment of the model in the following manner:

Taking the actual carbon content of molten steel as the standard, through repeated training and optimization selection, the parameters involved in the SVM learning algorithm are determined, which specifically include:

Construct characteristic parameters that can characterize the change of carbon content in the furnace from the flame spectrum information;

Select the kernel function of the SVM learning algorithm;

Optimize the control parameters kernel function width δ and penalty factor C;

Select the model training samples, and use the SVM learning algorithm to classify and model the characteristic parameters;

Input the established model with the test sample, and analyze whether the error and generalization meet the design requirements: if it is satisfied, then output the model, if not, return to the above steps to re-select the kernel function width δ and the penalty factor C to Remodel until requirements are met.

In a further embodiment, in the SVM carbon content dynamic prediction model, the kernel function of the SVM learning algorithm is selected from one of linear kernel function, polynomial kernel function, RBF kernel function and S-type kernel function.

In a further embodiment, in the SVM carbon content dynamic prediction model, after the computing unit receives the flame spectrum information collected in real time online and constructs characteristic parameters, it first determines the carbon content of the end point through the SVM carbon content dynamic prediction model. category, and based on the category of carbon, the corresponding endpoint fitting function is used to determine the carbon content of molten steel corresponding to the currently collected flame spectrum information.

In a further embodiment, the endpoint fitting function includes the endpoint fitting functions of different endpoint carbon categories, wherein:

The end point fitting function is expressed as:

Y=f(X),

This formula expresses the mapping relationship between X and Y, where X is the characteristic variable extracted from the flame spectrum at the end point, and Y is the carbon value at the end point. The end point fitting function uses a polynomial fitting function provided by MATLAB to fit the data. to get the fitting function.

It should be understood that all combinations of the concepts described, as well as additional concepts described in more detail below, may be considered part of the inventive subject matter of the present disclosure, provided such concepts are not mutually inconsistent. Additionally, all combinations of claimed subject matter are considered a part of the inventive subject matter of this disclosure.

These and other aspects, embodiments and features of the present teachings can be more fully understood from the following description when taken in conjunction with the accompanying drawings. Other additional aspects of the invention, such as the features and/or advantages of the exemplary embodiments, will be apparent from the description below, or learned by practice of specific embodiments in accordance with the teachings of the invention.

Description of drawings

The figures are not intended to be drawn to scale. In the drawings, each identical or nearly identical component that is illustrated in various figures may be represented by a like reference numeral. For purposes of clarity, not every component may be labeled in every drawing. Embodiments of the various aspects of the invention will now be described by way of example with reference to the accompanying drawings, in which:

Fig. 1 is a schematic diagram of the online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking based on SVM proposed by the present invention.

FIG. 2 is a schematic diagram of an endpoint control device based on SVM in the detection system of FIG. 1 .

FIG. 3 is a schematic diagram of the telescopic optical system in the detection system of FIG. 1 .

Fig. 4 is a schematic flow chart of the SVM classification modeling proposed by the present invention.

Fig. 5 is a schematic diagram of the data selection range proposed by the present invention.

6a-6b are schematic diagrams of classification hit results of training samples and test samples when penalty factor C=30 and kernel function width δ=0.2.

7a-8b are schematic diagrams of classification hit results of training samples and test samples when penalty factor C=20 and kernel function width δ=0.8.

8a-8b are schematic diagrams of classification hit results of training samples and test samples when penalty factor C=20 and kernel function width δ=3.

9a-9b are schematic diagrams of classification hit results of training samples and test samples when penalty factor C=100 and kernel function width δ=0.06.

Figure 10 is a schematic diagram of the classification model results of 30 carbon training.

FIG. 11 is a schematic diagram of the result of the final trained classification model.

FIG. 12 is a schematic diagram of the workflow of the detection system shown in FIG. 1 after actual startup according to the present invention.

Fig. 13 is a schematic diagram of the carbon content detection process proposed by the present invention.

detailed description

In order to better understand the technical content of the present invention, specific embodiments are given together with the attached drawings for description as follows.

Aspects of the invention are described in this disclosure with reference to the accompanying drawings, which show a number of illustrated embodiments. Embodiments of the present disclosure are not necessarily intended to include all aspects of the invention. It should be appreciated that the various concepts and embodiments described above, as well as those described in more detail below, can be implemented in any of numerous ways, since the concepts and embodiments disclosed herein are not limited to any implementation. In addition, some aspects of the present disclosure may be used alone or in any suitable combination with other aspects of the present disclosure.

The online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking proposed by the present invention, generally speaking, obtains the flame image information of the furnace mouth from a long distance through the telescopic system, and uses a spectrometer to analyze it to obtain the flame spectrum information. Therefore, useful information that can express carbon content is extracted from it to construct characteristic parameters, and SVM training is performed on the selected training samples to obtain a detection model. On-line real-time detection of carbon content in molten steel.

Since the current converter steelmaking generally carries out the corresponding converter blowing operation according to the different steel types, the steel types to be blown according to the requirements are divided into low-carbon steel (the content of carbon in molten steel is less than 1.5%), medium-carbon steel There are three kinds of steel (the content of carbon in molten steel is 1.5%～3.0%) and high-carbon steel (the content of carbon in molten steel is greater than 3.0%). The 15 carbons mentioned in the disclosure of the present invention are actually per thousand The 15th, that is, the carbon content is 1.5%, and the same is true for the 30th carbon. The 30th carbon model described below is the model obtained by performing classification training near several converters with a carbon content of 3.0%.

As shown in FIG. 1 , the SVM-based online real-time dynamic detection system for the carbon content of molten steel in converter steelmaking proposed by the present invention includes a telescopic optical system 1 , a spectrometer 2 and an SVM-based endpoint control device 3 .

The telescopic optical system 1 is connected with the spectrometer 2 through an optical fiber 4 .

The telescopic optical system 1 is configured to collect flame image information of a steelmaking furnace mouth in real time.

The spectrometer 2 is configured to receive the flame image information from the telescopic optical system 1 through the optical fiber 4, and obtain the flame spectrum information of the flame image information.

Spectrometer 2. In this example, a grating spectrometer is selected, such as Ocean Optics’ USB4000-VIS-NIR miniature CCD grating spectrometer, which is small in size, low in failure rate, and easy to install. It can be used in conjunction with the telescopic optical system designed in this example. The stable spectrum of the flame at the furnace mouth can be obtained stably.

Based on the SVM endpoint control device 3, the device has a computing unit 31 and a central control unit 32 that controls the operation of the computing unit. Real-time detection of carbon content in molten steel for steelmaking.

In this example, the SVM-based endpoint control device 3 is configured as a circuit board. As shown in Figure 2, the circuit board 3 is integrated with an FPGA chip as an operation unit and a microprocessor as a central processing unit. Of course, the circuit board 3 also includes a power supply module for providing a stable voltage supply, a serial interface, RS232 interface and other interfaces.

The SVM carbon content dynamic prediction model is burned in the FPGA chip, and the carbon content is automatically detected after receiving the flame spectrum information.

Of course, in an alternative implementation manner, the arithmetic unit may also be realized by a CPLD chip.

3, in the detection system of the present disclosure, the telescopic optical system 1 includes an objective lens with a common optical axis, an eyepiece, and a field of view diaphragm independent of the objective lens and the eyepiece. The formed optical imaging channel is used to adjust the field of view of the flame detection at the furnace mouth.

As an optional solution, the aforementioned field stop is located between the objective lens and the eyepiece, preferably on the focal plane of the objective lens.

In another example, the field stop is located behind the eyepiece and adjacent to the optical fiber.

As an optional example, the objective lens is a double split lens, which is composed of a positive lens and a negative lens with a common optical axis.

The eyepiece is a Kenel eyepiece, which is composed of a single lens and a doublet lens with a common optical axis distribution.

Preferably, the field of view diaphragm is a variable field of view diaphragm.

As shown in Figure 3, the symbol l represents the optical axis of the objective lens and the eyepiece, f ₁ ' represents the focal length of the objective lens, and f ₂ ' represents the focal length of the eyepiece.

As previously mentioned, in the end control device 3 based on SVM, the computing unit 31 is realized by one of FPGA and CPLD, and the SVM carbon content dynamic prediction model is burned in FPGA or CPLD, and when receiving the flame spectrum After the information is controlled by the central control unit 32, the detection of the carbon content is carried out automatically.

In this example, the SVM carbon content dynamic prediction model includes:

A parameter building block for constructing a characteristic parameter characterizing the change of carbon content in the furnace according to the input flame spectrum information;

A carbon content dynamic prediction module for performing carbon content prediction based on the constructed characteristic parameters; and

Output module for predicting result output.

In the SVM carbon content dynamic prediction model, the carbon content dynamic prediction module for performing carbon content prediction based on the constructed characteristic parameters, in this example, preferably realizes the establishment of the model in the following manner:

Taking the actual carbon content of molten steel as the standard, through repeated training and optimization selection, the parameters involved in the SVM learning algorithm are determined, which specifically include:

Construct characteristic parameters that can characterize the change of carbon content in the furnace from the flame spectrum information;

Select the kernel function of the SVM learning algorithm;

Optimize the control parameters kernel function width δ and penalty factor C;

Select the model training samples, and use the SVM learning algorithm to classify and model the characteristic parameters;

Input the established model with the test sample, and analyze whether the error and generalization meet the design requirements: if it is satisfied, then output the model, if not, return to the above steps to re-select the kernel function width δ and the penalty factor C to Remodel until requirements are met.

When the actual model is generated, the carbon value span of the data is narrowed by dividing the training to be tested into three categories.

As mentioned above, when performing model training and establishment, it adopts the method of classification modeling.

In this example, as shown in Figure 4, the specific implementation of classification modeling includes:

1) Selection of training and testing samples

Collect 100 flame spectrum data of the flame at the mouth of the converter from the steel mill, classify the 100 spectrum data, and divide them into three categories with carbon content C<15, 15≤C<30, and C≥30, and classify them from SVM The algorithm knows that for multi-classification problems, it can be transformed into several binary classification problems, which simplifies the difficulty of the algorithm and reduces the time complexity of multi-classification. For the present invention, it is obviously a multi-classification problem, so two binary classification models can be used to solve it. During training, only two models need to be trained near 30 carbons and 15 carbons to achieve the goal of dividing the data into three categories. Purpose.

Exemplarily, assuming that the range of carbon is between 0 and 50, the selection of specific training data is shown in FIG. 5 .

Therefore, for the training of the 30-carbon classification model, the ranges of its two classes are: between 0 and 30 carbons and above 30 carbons (the measured maximum value is around 50 carbons). After determining the two classes, determine the samples used for training and testing. The number of training samples and test samples is generally selected according to the ratio of 2:1 or 3:1. It is best not to have the same number, which is not conducive to the model. Generalization, the number of training samples is better than the number of testing samples.

Similarly, for the training of the 15-carbon model, its two categories are: between 15 and 30 carbons and below 15 carbons, and the number of training and testing samples is the same as the selection rule for the 30-carbon model.

After the sample is determined here, it is not static. It is possible that the selected sample will not meet the requirements no matter how it is done during the training of the classification model. At this time, it is necessary to change the sample, and put the unhit furnaces in the test sample into the training sample for retraining. Here, it needs to be explained that if the training sample is changed at this moment, then the next moment of model training, the sample will follow. Change, that is to say, the training samples at different times are consistent.

2) Construction of feature variables

The light intensity at the furnace mouth can represent the brightness of the flame, and the spectral information of the furnace mouth flame can represent the state in the converter. In the process of converter steelmaking, the state of molten steel in the furnace always changes from one state to the next, so it can be represented by a quantified value, that is, Yd can be used to represent the state information in the converter. This value includes Furnace temperature and composition information. Here Yd can be discrete or continuous. When observing the change trend of the flame spectrum, no abnormal mutation of the spectrum was found, but a slight change in certain bands of the spectrum was found. This change will be more obvious at the end of blowing, and may become a sign of the end point judgment .

Therefore, in this example, M(λ) is used to represent the spectral distribution of the flame at the furnace mouth, and Yd is considered to be a continuously changing value, then a mapping can be used to convert the independent variable (that is, the value of M(λ)) into a dependent variable (ie Yd value), the following functional relational formula is obtained: f(M(λ))=Yd.

The state of molten steel in the blowing process is related to the spectrum of the flame at the furnace mouth, that is, the state of molten steel in the furnace can be predicted through the distribution of the spectrum. Therefore, it is necessary to preprocess the spectral distribution, ignoring the influence of the absolute value of the spectral intensity, and only considering the influence of the change of the spectral distribution on the state of molten steel.

For the spectral data at a certain moment, the following formula can be used to process:

${\mathrm{<span\; class="notranslate">\; m</span>}}^{<span\; class="notranslate">\; \&prime;</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}{{\mathrm{<span\; class="notranslate">\; max</span>}}_{\mathrm{<span\; class="notranslate">\; 350</span>}<span\; class="notranslate">\; \&le;</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; 1000</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; \&lambda;</span>}}_{\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span>}$

In the above formula, M(λ _i ) represents the absolute value of the spectral light intensity at the wavelength λ _i , and its function is to normalize the spectral light intensity. After processing, the shape of the spectral distribution does not change, but Its value is between 0 and 1 during the blowing process.

The value range of the wavelength λ _i depends on the detection range of the spectrometer 2. In this example, the detection range of the spectrometer used is [350, 1000].

During the research process, there are two obvious raised peaks in the flame spectral distribution, and they change most regularly throughout the blowing process in many studies, especially when the end of the blowing is close to the end, their shape will change To a certain extent, it represents the state of molten steel in the furnace. The corresponding wavelengths are 600nm, 770nm, and 772nm respectively. Therefore, in the present invention, the normalized values of the spectra at these two places are used as characteristic variables a ₁ and a ₂ . That is, a ₁ =M'(600), since the peak at 700nm is a double peak, it is averaged, that is, the normalized mean value of light intensity at 770nm and 772nm is calculated:

In the process of many researches, it is also found that the continuous spectral change between the two peaks is also relatively sharp, which can reflect the change of flame brightness. In order to obtain the characteristic variable that can characterize this section of the spectral line, we will use this in this example A spectral line is divided into three sections on average, and the average value of the normalized light intensity of each section is taken to obtain three characteristic parameters: a ₃ , a ₄ , and a ₅ .

The peak of the spectral distribution can also reflect certain information, but the normalized value of the light intensity of the peak has been reflected in the parameter a _4. When selecting parameters, the first thing to ensure is that the parameters are independent of each other and cannot exist among them. Two parameters can reflect another parameter, or the situation that the parameters reflect each other, so there is no need to repeatedly select.

For this case, in this example, the ratio of the peak wavelength of the spectrum in the spectral distribution to the maximum value of the detection range of the spectrometer is used as the sixth parameter: a ₆ .

Therefore, in this embodiment, the parameter construction module is configured to construct characteristic parameters in the following manner:

The shape of the spectrum at a wavelength of 600nm is a raised peak, and the characteristic parameter a ₁ is the normalized value of the light intensity here;

The peak of the spectral shape raised at 770nm is a _double peak, and the characteristic parameter a2 is the normalized mean value of the light intensity at the wavelength of 770nm and 772nm;

The continuous spectrum in the middle of the two sharp peaks changes drastically, and the spectral line is divided into three segments on average, and the light intensity of each segment is normalized and averaged to obtain three characteristic parameters a ₃ , a ₄ , and a ₅ ;

The ratio of the peak wavelength of the spectrum in the spectral distribution to the maximum value of the detection range of the spectrometer is used as the sixth parameter: a ₆ .

Through the above analysis and introduction, the value ranges of the six parameters are all between [0,1], this is done to maintain the robustness of the algorithm. The spectral parameters have been selected here, and the parameters can be obtained by spectral calculation. For N training samples, each training sample contains many data points x _i =(a ₁ ,a ₂ ,a ₃ ,a ₄ ,a ₅ , a ₆ ), i=1,...,N, each data point contains 6 parameters, and then the next step is to import sample parameters and use SVM to train the classification model.

3) Selection of training parameters

When classifying and training samples here, the selection of samples must follow a condition: the samples are independent and identically distributed. After analysis, the spectral information at each moment reflects the state in the furnace. For the converter blowing process, each state is independent of each other, that is to say, the flame spectral information at each moment is also independent of each other. There is a corresponding relationship between each sample point, and then it can be obtained that each sample point is independent. And because the spectral distribution at each moment is obtained by burning the same medium, they must satisfy the same physical characteristics, so they conform to the same probability distribution. In summary, it can be considered that the selected training samples are independent and identically distributed.

Different kernel functions correspond to different grid structures during SVM training. In this example, the kernel function of the SVM learning algorithm is selected from one of linear kernel function, polynomial kernel function, RBF kernel function and S-type kernel function.

Linear kernel function: K(x,x')=<x,x'>;

Polyphase kernel function: K(x,x')=(<x,x'>+1) ^d where d is a positive real number;

S-type (sigmoid) kernel function: K(x,x')=thanh(v<x,x'>+r) where v and r are positive constants.

Gaussian (radial basis function, RBF) kernel function: K(x,x')=exp(-||x-x'|| ² /2σ ² ), where σ is the kernel width and is a positive integer. (Reference: A Practical Guide to Support Vector Classification page 2, Hsu CW, Chang CC, Lin C JA practical guide to support vector classification[J].2003.)

In practical applications, it is usually necessary to select the appropriate kernel function and parameters according to the specific situation of the problem. In order to find the most suitable SVM model kernel function corresponding to different converters, the above four kinds of kernel function training can be used for the sample spectral data of different converters, and then the kernel function with the best training result is selected as the converter SVM model kernel function

SVM classification is to find the hyperplane with the largest interval in the kernel function space after the kernel function is selected.

In this example, the LS_SVM algorithm is used to train the classification model, and the optimization definition of the original problem is obtained as:

$\mathrm{<span\; class="notranslate">\; Minimize</span>}<span\; class="notranslate">\; :</span><span\; class="notranslate">\; \&lang;</span>\mathrm{<span\; class="notranslate">\; \&omega;</span>}<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; \&omega;</span>}<span\; class="notranslate">\; \&rang;</span><span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; C</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; l</span>}}{\mathrm{<span\; class="notranslate">\; \&xi;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}$

Subject to: y _i (<ω·x _i >+b)≥1-ξ _i , i=1,...,l [Remark: Introduction to Support Vector Machines Page 91] In the above formula, C needs to be given The value of , which is a penalty factor for sample error.

The next job is to train the model. The specific steps are: import the spectral information of the training samples in chronological order, that is, the spectral parameters of each furnace. The training rule selected here is to train from the end point forward, that is, select the last 100 frames to train a model, and then continue to push forward 100 frames to train a model.

Firstly, train the furnaces within a certain period of time. For example, for the training samples, select 100 frames of data between the last 200 frames and 300 frames to train the model. Here, the extraction of 100 frames of data is consistent for all training samples.

Taking the 30-carbon classification model as an example, after the kernel function is determined, the next control parameters are mainly: kernel function width δ and penalty factor C.

The parameter adjustment methods mainly include: intelligent particle swarm optimization (PSO) algorithm (such as Eberhart R C, Kennedy J.A new optimizer using particle swarm theory[C]//Proceedings of the sixth international symposium on micro machine and human science.1995,1:39-43. ), genetic (GA) algorithm (such as Lei Jian. Modeling and global optimization method based on SVM and genetic algorithm [J]. Science and Technology Square, 2008 (5): 120-122.), grid search method (such as Wang Jianfeng , Zhang Lei, Chen Guoxing, etc. SVM parameter optimization based on the improved grid search method [J]. Applied Science and Technology, 2012, 39(3): 28-31.), etc., can use the above different methods to adjust parameters, combined with The hit rate of the sample and the generalization of the model, select the optimal kernel function width δ and penalty factor C.

The following still uses the 30-carbon classification model as an example to briefly introduce the parameter adjustment process:

First, randomly select a pair of penalty factor C = 30 and kernel function width δ = 0.2 parameters, and use SVM training to obtain a classifier, that is, a hyperplane. Its own training accuracy and test sample accuracy are shown in Figures 6a and 6b , the asterisk indicates the actual carbon value, and the circle indicates the missed heat, that is, the misclassified heat.

As shown in Figures 6a and 6b, for the classification hits of training samples and test samples, it can be seen from Figures 6a and 6b that the model has a very good classification effect on training samples, and all of them can be classified correctly. Under this parameter, all training samples are support vector. But for the test sample, there are 4 heats of carbon that are not classified correctly. If the other parameters are not considered, this situation is considered acceptable, but in reality it is necessary to try a large number of parameters to find a suitable model. As mentioned earlier, when selecting a model, the main considerations are the hit rate of the sample and the generalization of the model, and the generalization can be seen from the classification of the test data.

Next, consider the classification situation under the condition of parameters C=20 and δ=0.8, the hit rates of training and testing samples are shown in Fig. 7a and 7b.

From Figure 6a, 6b, Figure 7a, 7b, it can be seen that under these two sets of parameters, the hit rate of training samples and test samples is the same, but in the case of parameters C = 20 and δ = 0.8, the hit rate of training samples The classification accuracy is reduced, the number of sample support vectors is reduced, the generalization of the model is better than the former, and the accuracy of the new test data is better than the former. So in this case, it is more appropriate to choose the model trained by the latter pair of parameters.

As mentioned earlier, the error of the hyperplane with the largest interval in N random samples S is not greater than:

${\mathrm{<span\; class="notranslate">\; err</span>}}_{\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; f</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&le;</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; N</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; d</span>}\mathrm{<span\; class="notranslate">\; ln</span>}\frac{\mathrm{<span\; class="notranslate">\; n</span>}}{\mathrm{<span\; class="notranslate">\; d</span>}}<span\; class="notranslate">\; +</span>\mathrm{<span\; class="notranslate">\; ln</span>}\frac{\mathrm{<span\; class="notranslate">\; N</span>}}{\mathrm{<span\; class="notranslate">\; \&delta;</span>}}<span\; class="notranslate">\; )</span>$

In the above formula, d=#sv represents the number of support vectors, and the above formula shows that the fewer the number of support vectors, the stronger its generalization ability. Through the analysis, it is found that adjusting the value of δ can change the number of support vectors, thereby changing the generalization of the classification model. Under certain circumstances, the larger the value of δ, the smaller the number of support vectors and the better the generalization. It will be more obvious when SVM regression fitting. But it is not that the bigger δ is, the better. When δ increases to a certain extent, the generalization of the model has not been improved, and it may become worse, and even the hit rate of training samples may also become worse, as shown in Figure 8a, 8b shows the classification hit results of training samples and test samples when parameters C=20 and δ=3.

It can be seen from Figures 8a and 8b that it is necessary to adjust the value of the penalty factor C at this time, and the value of C represents the penalty factor for exceeding the target. Although the number of previous support vectors is reduced, the constraint ability of the parameter C is reduced relative to the training vectors, causing the training results that are not support vectors to exceed a certain range of the generalization error boundary, thus affecting its hit rate. So these two parameters need to be adjusted, not one parameter is fixed, just change the other parameter to find a suitable classification model, but both parameters need to be changed to select the classification model. By trying a large number of parameters, the classification model under the parameter combination of C=100 and δ=0.06 was finally selected, and the hit rate used to classify samples is shown in Figures 9a and 9b.

So far, for the classification model of 30 carbons, the classification model between the bottom 100 and 200 frames has been trained. According to the same optimization principle, 15 classification models up to the bottom 1500 frames can be found, and all model training Condition. When the flame spectral data is actually detected, the furnace mouth flame spectral data collected from the spectrometer first enters the 15th model, and then enters the 14th, 13th,..., 2nd, 1st classification models in turn, as shown in Figure 10. A label value will be obtained for each classification model, and this result value indicates the classification of the model, that is, which class it belongs to.

What needs to be clarified here is that the detection results obtained after the spectral data collected in the early stage of blowing enter the model are inaccurate, because the classification model of the present invention is trained according to the spectral data in the later stage of blowing, and only converter blowing enters the model. After the later stage, the classification model began to really play a role, which is consistent with the experience control of steel mill workers. The workers also started the real end point control at the end of blowing, and the early and middle stages were only based on long-term experience. .

Similarly, the same is true for the 15-carbon classification model, which is trained sequentially based on 100 frames. After the 30-carbon classification model and the 15-carbon classification model are all trained, the model detection process of the actual spectrum can be expressed as shown in Figure 11 It shows that there will be two classification models working at each time period. After the model is tested, it will know exactly which category the carbon content belongs to among high, medium and low, and then use the known fitting curve to detect actual carbon value.

As shown in Figure 12, after the carbon content online real-time dynamic detection system proposed by the present invention is started, when the on-site test is performed, the trained classification model will also start to classify, and it will give a classification result in real time. Indicates which category the carbon at this time belongs to: high, medium, or low. Obviously, after determining the carbon category, the next work process is to obtain the actual carbon value at this time, so it can be realized only by combining the end point fitting curve.

As shown in Figure 12, in the detection system proposed by the present invention, the SVM-based endpoint control device 3, such as a prediction board, can be connected to an upper industrial computer through a data line to receive and send data information to realize Debugging and control of the entire detection system, uploading, displaying, storing and subsequent analysis of data, such as displaying the spectral information collected in real time online on the display screen of the industrial computer, and/or displaying the predicted carbon content results in real time Displayed or graphically represented to the operator via the display.

In this example, that is to say, after the computing unit receives the flame spectrum information collected in real time online and constructs the characteristic parameters, it first determines the category of the end carbon through the SVM carbon content dynamic prediction model, and based on the category of the end carbon The carbon content of the molten steel corresponding to the currently collected flame spectrum information is determined by using the corresponding end point fitting function.

Certainly, the end point fitting function includes end point fitting functions of different types of end point carbons (eg low carbon steel, medium carbon steel, high carbon steel). That is to say, the furnaces are divided into three categories according to the carbon content: high, medium and low, and the samples of each category are fitted separately to obtain a fitting function, that is, three fitting curves of high, medium and low will be obtained in the end .

The end point fitting function Y=f(X) actually expresses the mapping relationship between X and Y, wherein X is the characteristic variable extracted from the flame spectrum at the end point, and Y is the carbon value of the end point. For example, use MATLAB to provide a polynomial fitting function to fit the data to obtain the fitting function.

An example of data fitting with a polynomial fitting function is given below.

The polynomial fitting function is as follows:

[p,s]=polyfit(X,Y,N)

The fitting criterion is the method of least squares, that is, to find such that The smallest f(x). N in the formula represents the order of fitting, p represents the coefficient vector of the polynomial, and s represents the error estimate for generating the predicted value.

The evaluation standard for selecting the fitting function is mainly the fitting accuracy, that is, the difference between the fitting value and the actual value within the error range. In order to consider the identification and observation of the fitted curve in practical applications, the downward trend of the fitted curve should also be considered.

After successfully training the required classification model and end-point carbon fitting function, as mentioned above, when the converter steelmaking reaches the end of blowing, the classification model starts to play its real role, and the carbon content in molten steel is monitored in real time. As shown in Figure 13, the exact carbon value is obtained by using the fitting function to provide real-time carbon content control basis for steelmaking plants.

Through the test, through the on-site detection of the furnace mouth flame spectrum data of 40 converters, the detection result of the end point carbon is obtained, and the hit situation and error distribution are analyzed, and the hit rate of the end point carbon classification detection method researched by the present invention can reach more than 85%. Fully meet the actual needs of steel mills.

Although the present invention has been disclosed above with preferred embodiments, it is not intended to limit the present invention. Those skilled in the art of the present invention can make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be defined by the claims.

Claims (10)

1. converter steel-smelting molten steel carbon content based on a SVM online Real-time and Dynamic Detection system, it is characterised in that this detection System includes:

Telescope optical system, is arranged to the flame image information of Real-time Collection steel-making fire door；

Spectrogrph, is configured to receive the flame image information from described telescope optical system by optical fiber, and obtains flame image The flame spectrum information of information；

Terminal point control device based on SVM, this device has the central authorities of an arithmetic element and control arithmetic element operation and controls single Unit, this arithmetic element is arranged for the flame spectrum information according to described real-time acquisition by SVM carbon content dynamic prediction mould Type carries out the real-time detection of carbon content in steel-making molten steel；

Wherein:

Described telescope optical system includes the object lens of common optical axis, eyepiece and independent of object lens and the field stop of eyepiece, this visual field In the optical imagery path that light hurdle is arranged in described object lens, eyepiece is formed, for regulating the visual field of fire door flame detecting.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 1 online Real-time and Dynamic Detection system, its Being characterised by, described field stop is positioned on the focal plane of described object lens.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 1 and 2 online Real-time and Dynamic Detection system, It is characterized in that, described field stop is variable field of view light hurdle.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 1 and 2 online Real-time and Dynamic Detection system, It is characterized in that, described field stop is positioned at described eyepiece rear and presses close to described optical fiber.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 1 online Real-time and Dynamic Detection system, its Being characterised by, described terminal point control device based on SVM, arithmetic element therein is real by the one in FPGA, CPLD Existing, described SVM carbon content dynamic prediction model burning is in described FPGA or CPLD, and is receiving flame spectrum Automatically the detection of carbon content is carried out after information.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 5 online Real-time and Dynamic Detection system, its Being characterised by, described SVM carbon content dynamic prediction model includes:

In in the flame spectrum information according to input, structure characterizes stove, the parameter of the characteristic parameter of carbon content change builds module；

The carbon content dynamic prediction module of carbon content prediction is carried out for characteristic parameter based on described structure；And

Output module for the output that predicts the outcome；

Wherein, described parameter builds module and is configured to construction feature parameter in the following manner:

At wavelength 600nm, spectral shape is protruding spike, characteristic parameter a₁For light intensity normalized value herein；

Spectral shape spike of projection at 770nm is bimodal, characteristic parameter a₂For the light intensity at wavelength 770nm and 772nm Normalization average value；

This section of spectral line acutely, is divided into three sections, to each section of light intensity normalizing by the continuous spectrum change in the middle of said two spike Average after change and obtain three characteristic parameter a₃, a₄, a₅；

Using the ratio of the peak wavelength λ of spectrum in spectral distribution and investigative range maximum of T max of described spectrogrph as the 6th Individual parameter: a₆。

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 6 online Real-time and Dynamic Detection system, its It is characterised by, in described SVM carbon content dynamic prediction model, described carry out carbon for characteristic parameter based on described structure and contain The carbon content dynamic prediction module of amount prediction, the foundation of implementation model in the following manner:

Using actual carbon content of molten steel as standard, by repetition training, optimized choice, determine involved by SVM learning algorithm Each parameter, it specifically includes:

The characteristic parameter of carbon content change in stove can be characterized by flame spectrum information architecture；

The kernel function of selected SVM learning algorithm；

Optimal control parameter kernel function width δ and penalty factor；

Selection Model training sample, utilizes SVM learning algorithm that characteristic parameter is carried out classification model construction；

Input the model set up with test sample, and whether analytical error and generalization meet design and require: if it is satisfied, then Output model, if be unsatisfactory for, then returns described step and re-starts the selection of kernel function width δ and penalty factor with again Modeling, until meeting requirement.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 7 online Real-time and Dynamic Detection system, its Being characterised by, in described SVM carbon content dynamic prediction model, the kernel function of described SVM learning algorithm is selected from linear kernel letter Number, Polynomial kernel function, the one in RBF kernel function and S type kernel function.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 1 online Real-time and Dynamic Detection system, its Being characterised by, in described SVM carbon content dynamic prediction model, described arithmetic element is receiving the fire of online real time collecting After flame spectral information construction feature parameter, first pass through described SVM carbon content dynamic prediction model and determine the classification of aim carbon, And classification of based on aim carbon uses corresponding terminal fitting function to determine the carbon of molten steel corresponding to current gathered flame spectrum information Content.

Converter steel-smelting molten steel carbon content based on SVM the most according to claim 9 online Real-time and Dynamic Detection system, It is characterized in that, described terminal fitting function include the classification of different aim carbon each belonging to terminal fitting function, wherein:

Described terminal fitting function is expressed as:

Y=f (X),

This formula have expressed the mapping relations of X Yu Y, and the characteristic variable extracted during wherein X is terminal moment flame spectrum, Y is Terminal carbon value, this terminal fitting function uses MATLAB to provide a polynomial fit function to be fitted data, thus Obtain fitting function.