CN105956618A - Converter steelmaking blowing state recognition system and method based on image dynamic and static characteristics - Google Patents

Abstract

The invention relates to a converter steelmaking and blowing state recognition system and method based on image dynamic and static features, including an image acquisition module, a color image flame area segmentation module, a color flame image dynamic and static feature representation and description module, and an image feature-based blowing state recognition system and method. Refining status classification module and upper computer monitoring module. The original image frame obtained through the video input, first measures and retains the pixels similar to the flame to achieve segmentation; then extracts four types of image features: brightness feature, chromaticity feature, texture feature, and optical flow field dynamic texture feature, and finally through the established The generalized regression neural network classifies and recognizes the input features, so as to achieve the functions of distinguishing the blowing stage and predicting the end point. The invention can adapt to the transient steady-state problem of the flame, has the advantages of high recognition accuracy and low cost, can be used in steelmaking converters of different scales, improves production efficiency and reduces waste of raw materials.

Description

System and method for converter steelmaking blowing status recognition based on image dynamic and static features

technical field

The invention relates to a converter steelmaking and blowing state recognition system and method based on dynamic and static features of images, belonging to the technical field of metallurgy automation.

Background technique

The iron and steel industry is an important raw material industry that supports the development of the national economy, and my country is the largest steel producer in the world. In the total production of steel, the average share of converter steel production reaches 70%. End point control is a key operation in the later stage of the converter. It refers to controlling the carbon content and temperature of molten steel to meet the requirements of tapping. Accurate and real-time judgment of the end point of converter blowing has always been one of the difficult problems in the iron and steel industry. Production efficiency, reducing energy and raw material waste, and improving steel quality are of great significance.

In the main blowing stage of blowing to about 85%, the secondary blowing strategy is adjusted by measuring the carbon content and temperature value. The most common data detection method is to use the sub-lance detection and experience judgment. The sub-lance is immersed in the molten pool for temperature measurement And take carbon, or the master worker takes samples from the furnace according to visual inspection, and adjusts the amount of oxygen blowing and the amount of raw materials added to the converter according to the measured data. For the judgment of the end point in the later stage of blowing, some iron and steel enterprises use the photoelectric detection method to judge the end point by the change of infrared laser passing through the furnace gas, and the gas analysis method through the determination of the chemical composition of the furnace gas. The detection equipment used in these methods needs to be Working in a high temperature and corrosive environment for a long time, the gas calibration cycle is short, the sampling head is replaced frequently, the use and maintenance costs of the equipment are high, and it is difficult to promote the use in the steelmaking converter industry.

Sub-gun control has high detection accuracy, but it is generally used in converters above 120t, which is difficult to meet the current situation of small and medium-sized steel mills in my country, and continuous measurement cannot be realized. The intelligent endpoint judgment method uses the molten steel detection data as the input vector of the model, and uses the target molten steel composition, temperature and oxygen blowing amount as the output vector to establish an endpoint prediction system based on ELM, CBR, GMDH and other models. Most of the intelligent endpoint prediction methods use the steelmaking process The blowing data actually collected in the method has good real-time performance in principle, but the existing methods have problems such as increased cost or untimely data acquisition in order to obtain accurate data.

With the rapid development of digital image processing technology and the improvement of computer processing performance, the image recognition based on artificial fire inspection has been rapidly developed and applied in the judgment of the end point of the converter. For example, the existing methods based on gray level co-occurrence matrix, based on color mean method, etc. Existing methods have achieved certain results in describing the static characteristics of the flame, but they have ignored the dynamic information of the flame, and the experience of actually watching the fire also shows that the dynamic characteristics of the flame can reflect the intensity of different stages of blowing, and can be used as the carbon oxidation rate. reflection. Therefore, combining the static and dynamic characteristics of the flame to achieve accurate identification of the blowing end point has important practical value and significance for improving production efficiency and reducing raw material waste.

Contents of the invention

In order to overcome the inaccuracy caused by manually checking the fire to determine the end point, the number of supplementary blowing times, low efficiency, waste of raw materials, etc., the present invention provides a converter steelmaking and blowing state recognition system and method based on dynamic and static features of images, and avoids The use of flue gas analysis, photoelectric analysis and other equipment is expensive, and the traditional flame image recognition does not consider the dynamic characteristics and affects the recognition rate.

The present invention is realized through the following technical solutions: a converter steelmaking and blowing state recognition system based on image dynamic and static features, including an image acquisition module, a color image flame area segmentation module, a color flame image dynamic and static feature representation and description module, and an image-based Features blowing state classification module and host computer monitoring module:

The image acquisition module is used to capture the flame situation in real time, and store and transmit the captured video signal to the color image flame area segmentation module;

The color image flame area segmentation module is used to divide the collected video signal into separate color image frames, and segment the flame background part and the flame body part in the color image;

The dynamic and static feature representation and description module of the colored flame image is used to collect the luminance information, chromaticity information, texture information, and dynamic information of the optical flow field of the segmented flame body;

The blowing state classification module based on image features is used to use the various information collected by the color flame image dynamic and static feature representation and description module as input to establish a recognition model, and after setting the output state, collect a data sample set for the built recognition model. Training; and then judge the output state of the newly collected input content, and send the output state of the final stage of blowing to the monitoring module of the host computer;

The upper computer monitoring module is used to send out signal prompts.

The image acquisition module also includes a camera, a memory and a signal transmitter.

Another object of the present invention is to provide a method for recognizing the blowing state of converter steelmaking based on the dynamic and static features of the image, through the following steps:

(1) Install a color camera so that it is consistent with the position and angle of the artificial fire, and fix the camera's position, focal length and other parameters; the purpose of fixing the camera is to ensure that the captured color images have feature comparability between frames and dynamic The accuracy of feature description; the size of the image is set to be equivalent to the perception ability of the human eye, the image is too small to affect the effect of feature extraction, and the image is too large to affect the processing speed of the algorithm; the real-time video signal collected by the color camera passes through the signal acquisition card and the processor connected;

(2) The processor divides the collected video signal into separate color image frames, and divides the flame background and the flame body in the color image, that is, converts the RGB color space to a uniform L*a*b* space, and records red ( 255,0,0) is converted to R _Lab , yellow (255,255,0) is converted to Y _Lab , white (255,255,255) is converted to W _Lab , let P be a pixel in the converted image, and D be between pixels The Euclidean distance of , according to the formula Then the segmentation strategy is:

$\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; P</span>}& \begin{array}{cc}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; f</span>}& {\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; T</span>}\end{array}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; w</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; e</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

In the formula, _DPI refers to the color difference value between one pixel of the image after color space conversion and the reference point pixel; I refers to a pixel after the image is converted from RGB color to Lab color, and this pixel has three components, respectively is L, a, b; T is the threshold for controlling the similarity;

then The part <T is used as the flame body, and the other parts are segmented as the flame background;

(3) Use the flame body part obtained in step (2) to collect brightness information, chromaticity information, texture information, and optical flow field dynamic information:

① Collect brightness information B:

Let I(i, j) be the divided flame area pixels, then the brightness feature B=(∑I(i, j))/(Count), where B is the brightness feature, Count is the number of area pixels, i, j is the position of the pixel, that is, the coordinates in the image;

②Collection of chromaticity information S:

Taking the color third-order moment of each component in the RGB space can effectively reflect the oxidation order of the melt pool elements during the blowing period, and the third-order moment under the single component i is

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 3</span>}}$

${\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}$

In the formula, p _i,j is the probability of occurrence of a pixel with gray level j in the i-th color channel component of the color image, i is the i-th component representing the color image, j is the gray value representing the image, and si is the image The color moment feature of , μi is the relevant parameter;

③Collect texture information ASM:

Use the gray difference statistical method to describe the static texture complexity, let (x, y) be a point in the image, and the gray difference between the point (x+Δx, y+Δy) is g _Δ (x, y) ＝g(x,y)-g(x+Δx,y+Δy), set all possible values of the gray difference as m levels, let the point (x, y) move within the given flame image area, and accumulate g The number of times _Δ (x, y) takes different values, and the histogram of g _Δ (x, y) is made. From the histogram, the probability of g _Δ (x, y) taking a value is p _Δ (i);

The second-order moment ASM of the extracted angle direction is used to reflect the uniformity of the gray distribution of the image. If the gray value of similar pixels has a large difference, the larger the ASM value, the rougher the texture;

$\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Texture features are related to the degree of oxidation of chemical elements in the molten pool, and the more intense the combustion, the higher the texture roughness;

④Collect dynamic information of optical flow field Ent:

In order to describe the dynamic information of the flame in the blowing process, especially the stroboscopic characteristics of the flame in the later stage, it is first necessary to establish a description model of the dynamic change of the flame, and use the optical flow field to establish the flickering process. The basic assumptions of the optical flow calculation are small motion and constant brightness , get I(x, y, t)=I(x+dx, y+dy, t+dt), and connect the points with constant pixel brightness in a limited area between adjacent frames to construct an inter-frame light Flow chart F; F records the intensity of the flame in the continuous change process, and the characteristic representation and description of F can reflect information such as stroboscopic;

Among them, x, y refers to the coordinates of the point in the image, t is the time axis of the continuous image, this is because the video is composed of images on the continuous time axis, dx represents the displacement of the x-axis, and dy represents the displacement of the y-axis , dt represents the displacement of the t-axis, and I is the image frame when the time is t in the video stream;

For the inter-frame optical flow map, the gray-level co-occurrence matrix is used to describe its characteristics. During the calculation process, the directions are 0°, 45°, 90°, and 135°, and the step size is step=1 to adapt to the characteristics of the random micro-texture of the flame. M is the obtained gray level co-occurrence matrix, using entropy to describe its characteristics, and calculating is the dynamic information of the flame;

Among them, i, j are the coordinates of pixels in the image;

(4) Adopt the generalized regression neural network to establish the identification model, the existing feature vector V=[B, S, ASM, Ent] collected in step (3) is input, and the state number of blowing is output: "1" Represents the early stage, "2" represents the middle stage, and "3" represents the end stage; the established generalized regression neural network has four inputs and one output, the activation function of hidden layer neurons is a Gaussian function, and the output is a linear mapping function; the known flame ontology The luminance information, chromaticity information, texture information, and optical flow field dynamic information are listed corresponding to their output data, and a data sample set is established; the built recognition model is trained, and the output data in training are classified according to the status number;

(5) After the real-time video signal collected by step (1) is processed by steps (2) and (3), it is input as a feature vector into the recognition model after step (4) training, and its output is the recognition of blowing state , when the blowing state output number is 3, it means that the blowing has entered the final stage, and a signal is sent out.

The point (x+Δx, y+Δy) of the step (3) is a point around the value pixel (x, y), and the distance between this point and the value pixel (x, y) is Δx and Δy.

Beneficial effects of the present invention: the present invention aims at the deficiencies of existing converter blowing state recognition methods based on flame image recognition, and proposes a state recognition scheme that combines dynamic and static image features; it is especially suitable for converters below 120 tons, or due to limited use of flue gas , secondary gun and other test environments. The advantage of the present invention is that it extracts the dynamic characteristics of the flame and combines the static characteristics to identify the blowing state, can adapt to the short-term steady-state transient problem of the flame, has the advantages of high recognition accuracy and low cost, and can be used in different scales. Steelmaking converters to improve production efficiency and reduce waste of raw materials. The invention solves the problems of inaccurate determination of the end point in the existing converter steelmaking process, and the high cost of use and maintenance of methods such as flue gas analysis, photoelectric analysis, sub-lance analysis, etc., and at the same time solves the problem of using the existing flame image recognition method In the analysis of a single static image, there are problems of large feature changes and low accuracy.

Description of drawings

Fig. 1 is the structure schematic diagram of the converter steelmaking blowing state recognition system based on image dynamic and static features of the present invention;

Fig. 2 is a schematic flow chart of the converter steelmaking blowing state recognition method based on the dynamic and static features of the image according to the present invention.

specific implementation plan

Hereinafter, embodiments of the invention will be described in detail, examples of which are illustrated in the accompanying drawings. The embodiments described with reference to the drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

As shown in Figure 1, the converter steelmaking blowing state recognition system based on image dynamic and static features includes image acquisition module 1, color image flame area segmentation module 2, color flame image dynamic and static feature representation and description module 3, image feature based The blowing state classification module 4 and the host computer monitoring module 5:

Image acquisition module 1 is used for real-time shooting of flame situation, and the video signal of shooting is stored and transmitted to the color image flame area segmentation module; image acquisition module also includes camera, memory and signal transmitter;

The color image flame area segmentation module 2 is used to divide the video signal of collection into separate color image frames, and the flame background part and the flame body part in the color image are segmented;

The dynamic and static feature representation and description module 3 of the colored flame image is used to collect brightness information, chromaticity information, texture information, and dynamic information of the optical flow field of the flame body part obtained from the segmentation;

The blowing state classification module 4 based on image features is used to use all kinds of information collected by the color flame image dynamic and static feature representation and description module as input to establish a recognition model, and after setting the output state, collect data sample sets for the built recognition model Carry out training; then judge the output state of the newly collected input content, and send the output state of the blowing into the final stage to the host computer monitoring module;

The host computer monitoring module 5 is used for sending out signal prompts.

As shown in Figure 2, the converter steelmaking blowing state recognition method based on the dynamic and static features of the image goes through the following steps:

(1) Install a color camera so that it is consistent with the position and angle of the artificial fire, and fix the camera's position, focal length and other parameters; the purpose of fixing the camera is to ensure that the captured color images have feature comparability between frames and dynamic The accuracy of feature description; the size of the image is set to be equivalent to the perception ability of the human eye, the image is too small to affect the effect of feature extraction, and the image is too large to affect the processing speed of the algorithm; the real-time video signal collected by the color camera passes through the signal acquisition card and the processor connected;

(2) The processor divides the collected video signal into separate color image frames, and divides the flame background and the flame body in the color image, that is, converts the RGB color space to a uniform L*a*b* space, and records red ( 255,0,0) is converted to R _Lab , yellow (255,255,0) is converted to Y _Lab , white (255,255,255) is converted to W _Lab , let P be a pixel in the converted image, and D be between pixels The Euclidean distance of , according to the formula Then the segmentation strategy is:

$\left\{\begin{array}{cc}\mathrm{<span\; class="notranslate">\; P</span>}& \begin{array}{cc}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; f</span>}& {\mathrm{<span\; class="notranslate">\; D.</span>}}_{\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; I</span>}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; Y</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; W</span>}}_{\mathrm{<span\; class="notranslate">\; L</span>}\mathrm{<span\; class="notranslate">\; a</span>}\mathrm{<span\; class="notranslate">\; b</span>}}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; <</span>\mathrm{<span\; class="notranslate">\; T</span>}\end{array}\\ \mathrm{<span\; class="notranslate">\; 0</span>}& \mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; t</span>}\mathrm{<span\; class="notranslate">\; h</span>}\mathrm{<span\; class="notranslate">\; e</span>}\mathrm{<span\; class="notranslate">\; r</span>}\mathrm{<span\; class="notranslate">\; w</span>}\mathrm{<span\; class="notranslate">\; i</span>}\mathrm{<span\; class="notranslate">\; the\; s</span>}\mathrm{<span\; class="notranslate">\; e</span>}\end{array}\right.<span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>$

In the formula, _DPI refers to the color difference value between one pixel of the image after color space conversion and the reference point pixel; I refers to a pixel after the image is converted from RGB color to Lab color, and this pixel has three components, respectively is L, a, b; T is the threshold for controlling the similarity;

then The part <T is used as the flame body, and the other parts are segmented as the flame background;

(3) Use the flame body part obtained in step (2) to collect brightness information, chromaticity information, texture information, and optical flow field dynamic information:

① Collect brightness information B:

Let I(i, j) be the divided flame area pixels, then the brightness feature B=(∑I(i, j))/(Count), where B is the brightness feature, Count is the number of area pixels, i, j is the position of the pixel, that is, the coordinates in the image;

②Collection of chromaticity information S:

Taking the color third-order moment of each component in the RGB space can effectively reflect the oxidation order of the melt pool elements during the blowing period, and the third-order moment under the single component i is

${\mathrm{<span\; class="notranslate">\; the\; s</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>{<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 3</span>}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; /</span>\mathrm{<span\; class="notranslate">\; 3</span>}}$

${\mathrm{<span\; class="notranslate">\; \&mu;</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}\underset{\mathrm{<span\; class="notranslate">\; j</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; C</span>}\mathrm{<span\; class="notranslate">\; o</span>}\mathrm{<span\; class="notranslate">\; u</span>}\mathrm{<span\; class="notranslate">\; no</span>}\mathrm{<span\; class="notranslate">\; t</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}$

In the formula, p _i,j is the probability of occurrence of a pixel with gray level j in the i-th color channel component of the color image, i is the i-th component representing the color image, j is the gray value representing the image, and si is the image The color moment feature of , μi is the relevant parameter;

③Collect texture information ASM:

Use the gray difference statistical method to describe the static texture complexity, let (x, y) be a point in the image, and the gray difference between the point (x+Δx, y+Δy) is g _Δ (x, y) ＝g(x,y)-g(x+Δx,y+Δy), set all possible values of the gray difference as m levels, let the point (x, y) move within the given flame image area, and accumulate g The number of times _Δ (x, y) takes different values, and the histogram of g _Δ (x, y) is made. From the histogram, the probability of g _Δ (x, y) taking a value is p _Δ (i); the point ( x+Δx,y+Δy) is a point around the value pixel (x,y) at distances Δx and Δy from the value pixel (x,y);

The second-order moment ASM of the extracted angle direction is used to reflect the uniformity of the gray distribution of the image. If the gray value of similar pixels has a large difference, the larger the ASM value, the rougher the texture;

$\mathrm{<span\; class="notranslate">\; A</span>}\mathrm{<span\; class="notranslate">\; S</span>}\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 0</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; \&Delta;</span>}}^{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; )</span>$

Texture features are related to the degree of oxidation of chemical elements in the molten pool, and the more intense the combustion, the higher the texture roughness;

④Collect dynamic information of optical flow field Ent:

In order to describe the dynamic information of the flame in the blowing process, especially the stroboscopic characteristics of the flame in the later stage, it is first necessary to establish a description model of the dynamic change of the flame, and use the optical flow field to establish the flickering process. The basic assumptions of the optical flow calculation are small motion and constant brightness , get I(x, y, t)=I(x+dx, y+dy, t+dt), and connect the points with constant pixel brightness in a limited area between adjacent frames to construct an inter-frame light Flow chart F; F records the intensity of the flame in the continuous change process, and the characteristic representation and description of F can reflect information such as stroboscopic;

Among them, x, y refers to the coordinates of the point in the image, t is the time axis of the continuous image, this is because the video is composed of images on the continuous time axis, dx represents the displacement of the x-axis, and dy represents the displacement of the y-axis , dt represents the displacement of the t-axis, and I is the image frame when the time is t in the video stream;

For the inter-frame optical flow map, the gray-level co-occurrence matrix is used to describe its characteristics. During the calculation process, the directions are 0°, 45°, 90°, and 135°, and the step size is step=1 to adapt to the characteristics of the random micro-texture of the flame. M is the obtained gray level co-occurrence matrix, using entropy to describe its characteristics, and calculating is the dynamic information of the flame;

Among them, i, j are the coordinates of pixels in the image;

(4) Adopt the generalized regression neural network to establish the identification model, the existing feature vector V=[B, S, ASM, Ent] collected in step (3) is input, and the state number of blowing is output: "1" Represents the early stage, "2" represents the middle stage, and "3" represents the end stage; the established generalized regression neural network has four inputs and one output, the activation function of hidden layer neurons is a Gaussian function, and the output is a linear mapping function; the known flame ontology The brightness information, chrominance information, texture information, and optical flow field dynamic information of the corresponding output data are listed, and the image and video signals of no less than 10 furnaces are established as the data sample set; and each image is blown Refine state labeling, train the built recognition model, and classify the output data in training according to state numbers;

(5) After the real-time video signal collected by step (1) is processed by steps (2) and (3), it is input as a feature vector into the recognition model after step (4) training, and its output is the recognition of blowing state , when the blowing state output number is 3, it means that the blowing has entered the final stage, and a signal is sent out.

Claims (4)

1. the pneumatic steelmaking blowing state recognition system moving static nature based on image, it is characterised in that include image acquisition Module, coloured image flame region split module, colored flames image sound state character representation and describing module, based on image spy The blowing state classification module levied and ipc monitor module:

Image capture module is used for captured in real-time flame conditions, and the video signal of shooting is stored and transmitted to coloured image fire Flame region segmentation module；

Coloured image flame region segmentation module is for being divided into single color image frames by the video signal of collection, to colour Flame background part in image is split with flame body part；

Colored flames image sound state character representation and describing module are for gathering the brightness letter of segmentation gained flame body part Breath, chrominance information, texture information and optical flow field multidate information；

Blowing state classification module based on characteristics of image is for by colored flames image sound state character representation and describing module The various information gathered is set up as input and is identified model, and after arranging output state, gathers the built knowledge of data sample set pair Other model is trained；And then the new input content gathered is judged output state, the output state that blowing enters latter stage is sent out Deliver to ipc monitor module；

Ipc monitor module is used for sending signal prompt.

2. according to the pneumatic steelmaking blowing state recognition system moving static nature based on image shown in claim 1, its feature It is: described image capture module also includes photographic head, memorizer and signal transmission device.

3. the pneumatic steelmaking blowing state identification method moving static nature based on image, it is characterised in that through following step Rapid:

(1) color video camera is installed so that it is keep consistent with the position manually seeing fire and angle, the position of fixing camera, Jiao Away from；The real time video signals that color video camera is gathered is connected with datatron by data acquisition card；

(2) video signal of collection is divided into single color image frames by datatron, to the flame background in coloured image with Flame body is split, and will change to uniform L*a*b* space by RGB color space, after note red (255,0,0) conversion For R_Lab, yellow (255,255,0) change after into Y_Lab, white (255,255,255) change after into W_LabIf P is converted images In a pixel, D is the Euclidean distance between pixel, according to formula Then segmentation strategy is:

$\left\{\begin{array}{cc}P& if{D}_{PI\left(I=R_{Lab},Y_{Lab},W_{Lab}\right)}<T\\ 0& otherwise\end{array}\right.---\left(1\right)$

In formula, D_PIRefer to the color value of chromatism between one pixel of the image after color space conversion and reference point pixel；I refers to An image pixel after RGB color is transformed into Lab color, this pixel has three components, is L respectively, a, b；T is for controlling The threshold value of similarity；

Again willPart as flame body, other parts are split as flame background；

(3) it is used for gathering monochrome information, chrominance information, texture information and optical flow field by step (2) gained flame body part Multidate information:

1. collection monochrome information B:

If (i, j) for the flame region pixel after segmentation, then (∑ I (i, j))/(Count), wherein B is bright to brightness B=to I Degree feature, Count is area pixel number, and i, j are the position of pixel, coordinate the most in the picture；

2. collection chrominance information S:

Third moment under simple component i is $s_i={\left(\frac{1}{Count}\underset{j=1}{\overset{Count}{\&Sigma;}}{\left(p_{i,j}-{\mathrm{\&mu;}}_i\right)}^3\right)}^{1/3}$

${\mathrm{\&mu;}}_i=\frac{1}{Count}\underset{j=1}{\overset{Count}{\&Sigma;}}{p}_{i,j}$

In formula, p_i,jThe probability occurred for the pixel that gray scale in coloured image i-th Color Channel component is j, i is for representing colour The i-th component of image, j is the gray value representing image, and si is the color moment feature of image, and μ i is relevant parameter；

3. collection texture information ASM:

Grey scale difference statistical method is utilized to describe static Texture complication, if (x is y) a bit in image, with point (x+ Δ x, y Gray scale difference value between+Δ y) is g_Δ(x, y)=g (and x, y)-g (x+ Δ x, y+ Δ y), if institute's likely value of grey scale difference For m level, make point (x, y) to flame image region in move, accumulative g_Δ(x y) takes the number of times of different value, makes g_Δ(x,y) Rectangular histogram, from rectangular histogram, g_Δ(x, y) probability of value is p_Δ(i)；

Use the uniformity coefficient extracting angle direction second moment ASM reflection gradation of image distribution, if close grey scale pixel value Differ greatly, then ASM value is the biggest, illustrates that texture is the most coarse；

$ASM=\underset{i=0}{\overset{m}{\&Sigma;}}{p}_{\mathrm{\&Delta;}}^2\left(i\right)$

Textural characteristics is relevant with the degree of oxidation of chemical element in molten bath, and the most violent grain roughness that burns is the highest；

4. collection optical flow field multidate information Ent:

Set up the descriptive model that flame dynamically changes, use optical flow field to set up scitillation process, according to the basic assumption of optical flow computation For small movements and brightness constancy, (x, y, t)=I (x+dx, y+dy, t+dt), by pixel intensity perseverance between consecutive frame to obtain I Fixed point is limiting region intraconnections, constructs interframe light flow graph F；F have recorded the flame violent degree in consecutive variations process, right F carries out character representation and description can reflect the information such as stroboscopic；

Wherein, x, y refer to the coordinate at image midpoint, and t is to be the time shaft of consecutive image, this is because video is by continuous time Image construction on axle, dx represents the displacement of x-axis, and dy represents the displacement of y-axis, and dt represents the displacement of t axle, and I is in video flowing It is the picture frame of t when the time；

To interframe light flow graph, using gray level co-occurrence matrixes to describe its feature, during it calculates, direction takes 0 °, 45 °, 90 °, 135 °, step-length step=1 is to adapt to the feature of the random microtexture of flame, if M is the gray level co-occurrence matrixes obtained, uses entropy Describe its feature, calculateMultidate information for flame；

Wherein, the coordinate of pixel during i, j are image；

(4) use generalized regression nerve networks to set up and identify model, existing characteristic vector V=that step (3) gathers [B, S, ASM, Ent] for inputting, with the state in which numeral that blows for output: " 1 " represents the initial stage, and " 2 " represent mid-term, and " 3 " represent latter stage；Build Vertical generalized regression nerve networks is four inputs and single output, and hidden layer neuron excitation function is Gaussian function, is output as line Property mapping function；By monochrome information, chrominance information, texture information and the optical flow field multidate information of known flame body and its Output data correspondence is listed, and sets up set of data samples；It is trained building identification model, the output data in training are pressed shape State numeral is sorted out；

(5) real time video signals step (1) gathered is after step (2) and (3) process, as characteristic vector input step (4) in the identification model after training, its output is the identification of blowing state, and when the State-output numeral that blows is 3, representative is blown Refining enters latter stage, sends signal prompt.

4. according to the pneumatic steelmaking blowing state identification method moving static nature based on image shown in claim 3, its feature Be: the point of described step (3) (x+ Δ x, y+ Δ y) be value pixel (x, y) point around, this point and value pixel (x, y) Distance be Δ x and Δ y.