Converter end point judgment method based on deep learning

Abstract

The invention relates to a converter end point determination method based on deep learning, and the method comprises the following steps: 1, obtaining the flame state of a complete smelting process through a camera in front of a flame at a converter mouth, and obtaining the input of a flame recognition system; 2, leading out frames from the input stream video, and adjusting a flame recognition region of interest of each frame of image, so that a flame feature region is more obvious and easy to recognize; 3, converting an RGB three-primary-color channel of the processed ROI region into an HSI channel, thereby deepening flame texture features, and completing a preprocessing process of a flame image; the method comprises the following steps: S1, carrying out Resnet101 network configuration, S2, carrying out Resnet101 network configuration, S4, carrying out data set division on an ROI region of a frame-by-frame HSI channel, S5, carrying out deployment on parameters and weights of the Resnet101 network, carrying out real-time processing on furnace mouth flames in a converter smelting process, inputting the processed furnace mouth flames into the network for detection, and returning a label value of a smelting period.

Description

Technical Field

[0001] The present invention relates to a determination method, and in particular to a converter endpoint determination method based on deep learning, and belongs to the technical field of automatic control. Background Art

[0002] Steel occupies an important position in the production activities of the national economy, and it indicates the degree of economic development of the country. BOF steelmaking has been widely used in my country's steelmaking industry because of its advantages of high production efficiency, good adaptability of raw materials, small investment, and quick results. The judgment of the end point of the converter blowing is an important operation at the end of the converter blowing. The prediction of the converter end point of my country's small and medium-sized steel mills mainly relies on manual experience and auxiliary gun detection to judge the converter end point. Manual experience relies on workers to visually observe the color, shape, texture, strobe, appearance, quantity and other characteristics of the converter flame and slag, and make a judgment on the converter end point based on experience. However, in actual situations, when the operating workers observe the converter flame, the work intensity is high and visual fatigue is easy to occur, especially during the night shift. In addition, there will be differences in the judgment of each individual worker, resulting in low accuracy of the end point judgment, and multiple furnace reversals are required for re-blowing, resulting in waste of raw materials and energy and affecting steelmaking efficiency. The auxiliary gun detection uses the auxiliary gun to contact and detect in the molten pool, detect the temperature, carbon content and other indicators in the molten steel, and judge the converter end point based on the collected data. The auxiliary gun detection is generally used in converters above 120t. Multiple detections will affect the service life of the probe, and real-time continuous measurement cannot be performed. The furnace gas analysis detection method uses relevant instruments to detect the furnace mouth composition, and obtains the corresponding indicators in the molten pool after analysis. This method can continuously measure the blowing data. Due to long-term use in a high-temperature corrosive environment, the use and maintenance costs of such equipment are high. The method based on intersection radiation uses the changes in the furnace mouth light radiation energy at different periods of converter blowing to judge the converter end point, but the multiple interference radiation sources at the converter site will affect the radiation collection instrument. Therefore, it is very valuable to study an accurate and real-time method for judging the converter steelmaking end point for improving the converter steelmaking process.

[0003] Patent document 112734722 discloses a method for judging the end point of converter smelting: first, the flame image is converted into color space and the texture information of each single channel is fully considered; then local phase quantization (LBQ) is used to extract image phase information as a supplement to the amplitude information in CLBP to form a fusion feature CLBP_MP to enhance the robustness of the CLBP algorithm structure, and the color information is weighted by an improved color information weighting strategy to enhance the local contrast information of the flame image; finally, the K nearest neighbor regression model is used to predict the carbon content. However, the invention loses the color feature of the flame after the flame is binarized, and only relies on abstract shape judgment, which has a low accuracy rate. Patent document CN111340116 discloses a converter flame recognition method, system, electronic device and medium, wherein the method comprises: collecting flame images and classifying them to obtain a data set; normalizing the image information in the data set; providing a neural network, wherein the neural network comprises a ResNet network; inputting the image information in the normalized data set into the neural network for training to obtain a training model; and determining the classification of the flame image by the training model. However, this method can only determine the converter smelting state, and is of no help in determining the converter smelting end point. Patent document CN113221827 discloses an intelligent converter flame state identification and re-inspection method, system, medium and terminal, the method comprising: obtaining converter flame image data, and dividing the abnormal state into several levels according to the degree of abnormality according to the splash type of the converter flame; establishing a converter flame recognition model and training it; collecting a real-time image of the converter flame, inputting the value into the trained converter flame recognition model, performing flame state recognition, obtaining the recognition result for synchronous display, and re-inspecting the recognition result according to the real-time image, but the main purpose of the invention is to judge whether the converter is splashing and the splashing level, and it is impossible to judge the temperature and carbon content of the molten steel in the converter through the converter flame, that is, it is impossible to judge the end point of the converter smelting. Therefore, a new solution is urgently needed to solve the above technical problems. Summary of the invention

[0004] The present invention aims at the problems existing in the prior art and provides a method for determining the converter endpoint based on deep learning. The technical solution realizes artificial intelligence judgment, which can reduce the risks brought by manual fire watching, reduce the influence of human subjective factors, and determine the converter endpoint more accurately.

[0005] In order to achieve the above object, the technical solution of the present invention is as follows: a method for determining a converter endpoint based on deep learning, characterized in that the method comprises the following steps:

[0006] Step 1: Obtain the flame state of the complete smelting process through the flame front camera of the converter mouth to obtain the input of the flame recognition system;

[0007] Step 2: By exporting frames of the input stream video and adjusting the flame recognition region of interest (ROI) of each frame image, the flame feature area is made more obvious and easy to identify;

[0008] Step 3, by transforming the RGB three primary color channels of the ROI area processed above into HSI channels, the flame texture features are deepened, and the preprocessing process of the flame image is completed;

[0009] Step 4: Divide the data set into the ROI area of ​​the HSI channel frame by frame, and put it into the Resnet101 convolutional neural network for training to obtain the weight file of converter flame detection;

[0010] Step 5: Deploy the parameters and weights of the Resnet101 network, process the furnace mouth flame in the converter smelting process in real time, input it into the network for detection, and return the label value of the smelting period.

[0011] As an improvement of the present invention, in step 2, the flame recognition region of interest (ROI) of each frame image is adjusted, and is selected through the OpenCV open source image library and according to the flame change law of different converter smelting periods.

[0012] As an improvement of the present invention, step 3 includes: extracting flame features from the original image frame, firstly through the selection of the ROI area, and then completing the conversion of the RGB channel of the flame to the HSI channel according to the following calculation formulas (1)-(4);

[0013]

[0014]

[0015]

[0016]

[0017] In the above formula, H represents the phase angle of color, which is the most important attribute of color and determines the nature of color; S represents the depth of color, the higher the saturation, the darker the color; I represents the brightness of the color. R represents the red channel; G represents the green channel; B represents the blue channel, and a multi-scale expression based on the texture and color characteristics of the converter flame is constructed.

[0018] As an improvement of the present invention, in step 4, for the establishment of the Resnet101 residual convolutional neural network, the Pytorch deep learning framework is used, the number of residual layers is 101, and the label files are "early", "mid", and "late" respectively; the tensor dimensions of the input of the training process are unified as [3,244,244].

[0019] As an improvement of the present invention, a converter endpoint determination system is used, the system comprising:

[0020] The acquisition module is used to obtain a continuous video stream of the flame state during the converter steelmaking process;

[0021] The processing module performs image processing on the continuous video stream, including the selection of the flame region of interest and the conversion of the RGB channel of the ROI region to the HSI channel;

[0022] The calculation module inputs the input image frame into the network model after preprocessing, performs forward calculation through the weight file, and returns the index value of the preset label;

[0023] The judgment module searches for the label and the corresponding confidence level through the preset status label index value to complete the judgment of the flame status at the smelting moment.

[0024] Compared with the prior art, the present invention has the following advantages: the technical solution realizes artificial intelligence judgment, can reduce the risk brought by manual fire watching, and can reduce the influence of human subjective factors. At the same time, compared with the direct use of the traditional Resnet network model, this method has a higher generalization ability for the complex multi-variable strong coupling and multi-state time-varying nature of the smelting process, and can judge the converter endpoint more accurately, which is conducive to the realization and integration of automatic converter steelmaking and one-click steelmaking. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Establish a flow chart for the converter flame state judgment method model;

[0026] Figure 2 This is a schematic diagram of the overall structure of the converter flame state judgment model. DETAILED DESCRIPTION

[0027] In order to deepen the understanding of the present invention, the present embodiment is described in detail below with reference to the accompanying drawings.

[0028] Embodiment 1: A method for determining converter endpoint based on deep learning, the method comprising the following steps:

[0029] Step 1: Obtain the flame state of the complete smelting process through the flame front camera of the converter mouth to obtain the input of the flame recognition system;

[0030] Step 2: By exporting frames of the input stream video and adjusting the flame recognition region of interest (ROI) of each frame image, the flame feature area is made more obvious and easy to identify;

[0031] Step 3, by transforming the RGB three primary color channels of the ROI area processed above into HSI channels, the flame texture features are deepened, and the preprocessing process of the flame image is completed;

[0032] Step 4: Divide the data set into the ROI area of ​​the HSI channel frame by frame, and put it into the Resnet101 convolutional neural network for training to obtain the weight file of converter flame detection;

[0033] Step 5: Deploy the parameters and weights of the Resnet101 network, process the furnace mouth flame in the converter smelting process in real time, input it into the network for detection, and return the label value of the smelting period.

[0034] Among them, in step 2, the flame recognition region of interest (ROI) of each frame image is adjusted, and the selection is made through the OpenCV open source image library and according to the flame change law of different converter smelting periods.

[0035] Step 3 includes: extracting flame features from the original image frame, firstly by selecting the ROI area, and then completing the conversion of the RGB channel of the flame to the HSI channel according to the following calculation formulas (1)-(4);

[0036]

[0037]

[0038]

[0039]

[0040] In the above formula, H represents the phase angle of color, which is the most important attribute of color and determines the nature of color; S represents the depth of color, the higher the saturation, the darker the color; I represents the brightness of the color. R represents the red channel; G represents the green channel; B represents the blue channel, and a multi-scale expression based on the texture and color characteristics of the converter flame is constructed.

[0041] In step 4, the Pytorch deep learning framework is used to build the Resnet101 residual convolutional neural network. The number of residual layers is 101, and the label files are "early", "mid", and "late" respectively. The tensor dimensions of the input in the training process are unified as [3,244,244].

[0042] In the above scheme, the flame state of the whole converter smelting process is obtained from the image acquisition device in front of the furnace; the flame features are enlarged from the original image by selecting the ROI region of interest to obtain a preliminary preprocessed image set; the network input image after complete preprocessing is obtained by converting the RGB channel to the HSI channel of the above ROI area, and the front, middle and back periods are divided; the final model weight file is obtained by putting the above established data set into the Resnet residual network for training; the model weight file and the network structure are deployed, and the converter flame image is processed in real time and forward calculated, so as to output the prediction of the current converter flame state. The flame video stream at the furnace mouth during the converter smelting process is obtained, including: according to the image acquisition device in front of the furnace, a complete round of converter slag discharge process is obtained to obtain a continuous video stream input for the converter furnace mouth. It is established through the pytorch deep learning framework, and the training and prediction effects of the residual convolutional networks of three different depths of Resnet34, Resnet50 and Resnet101 are compared, including parameters such as lr learning rate and loss loss. The network model selected for comparison is Resnet101, and the tracing method under the pytorch framework is used to convert and save the model. The converted model is in torchscript format. For model deployment, libtorch is used to load the converted weight file, the programming language is C++, and the interface development platform is Qt.

[0043] Embodiment 2: A converter endpoint determination system based on deep learning, the system comprising:

[0044] The acquisition module is used to obtain a continuous video stream of the flame state during the converter steelmaking process;

[0045] The processing module performs image processing on the continuous video stream, including the selection of the flame region of interest and the conversion of the RGB channel of the ROI region to the HSI channel;

[0046] The calculation module inputs the input image frame into the network model after preprocessing, performs forward calculation through the weight file, and returns the index value of the preset label;

[0047] The judgment module searches for the label and the corresponding confidence level through the preset status label index value to complete the judgment of the flame status at the smelting moment.

[0048] See also Figure 2 , Figure 2 The following is a simplified flow chart of establishing a converter flame detection program module in one embodiment of the present application, wherein the structure may include the following modules:

[0049] Acquisition and processing module: This module includes the acquisition of converter smelting flue gas data and the acquisition of furnace front flame image data. The preprocessing of the input image includes ROI area selection and HSI channel conversion. As the smelting progresses, the process flame is collected in real time through the camera and updated to the original training data set.

[0050] Calculation and prediction module: This module includes the training stage after the model data set is updated, and calculates by inputting the image into the model, and predicts the turning points of the second and third stages, and then inputs the smelting turning point information of the furnace, and inputs the results to the next stage.

[0051] Forecast module: accepts network forecast information, completes the countdown for the output smelting endpoint forecast, and highlights the forecast prompt on the on-site worker interface.

[0052] Example 3: See Figure 1 , Figure 1 A simplified flow chart of a converter flame state judgment method model establishment in an embodiment of the present application is shown, and the method may include steps S101-S105:

[0053] Step S101, the camera in front of the converter acquires the original video stream image of the converter smelting process,

[0054] Step S102, by selecting the region of interest ROI of the original image frame, amplifying the flame characteristics of the converter at different stages,

[0055] Step S103, converting the RGB image of the ROI area into an HSI channel to obtain a pre-processed image frame before inputting into the model.

[0056] Step S104, divide the data set according to the flue gas data of converter smelting, and input it into the Resnet101 network for training to obtain the output weight file.

[0057] Step S105, completing the deployment of the converter flame state judgment model through the weight file and network structure parameters trained by the above network.

[0058] In this application, the flame characteristics of the converter mouth were screened, and the areas with obvious changes in the flame characteristics in the early, middle and late stages were selected as the region of interest (ROI). The flame in the early stage is darker, mostly dark red; the flame in the middle stage is brighter, and the flame area is larger; the color difference in the late stage is not obvious compared to the middle stage, but the flame shape is contracted, and the flame area is smaller than the middle stage.

[0059] In this application, by converting the ROI area selected in the above process from RGB channel to HSI channel, the texture of the converter flame under the RGB channel is not obvious, which is not conducive to the observation of data set features. When converted to the HSI channel, the texture contrast of the bright yellow flame is more obvious and the features are richer, so that the data set can be divided more accurately.

[0060] In the present application, the method for judging the flame state of the converter mouth of the embodiment can be applied to electronic devices, such as computer devices. Specifically, the electronic device is connected to the image acquisition device, and the electronic device also has other modules, such as the image acquisition device and the electronic device are connected through a network, a USB cable or a serial port, and the electronic device also includes devices such as a processor and a memory. The image acquisition device can use a visible light camera or an infrared camera.

[0061] In this application, the flame image of the converter mouth is continuously acquired during the converter smelting process. The image acquisition device can be installed 5m in front of the converter mouth. The lens angle can fully record the entire smelting process. The acquired video stream is input into the image processing device. The image processing device can control the video stream to export the picture frame interval to 10ms, so that the flame at the current moment can be processed quickly and accurately to be input into the network for detection.

[0062] It should be noted that the above embodiments are not intended to limit the protection scope of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions all fall within the protection scope of the claims of the present invention.

Claims

1. A converter endpoint determination method based on deep learning, characterized in that: The method comprises the following steps: Step 1: Obtain the flame state of the complete smelting process through the flame front camera of the converter mouth to obtain the input of the flame recognition system; Step 2: By exporting frames of the input stream video and adjusting the flame recognition region of interest (ROI) of each frame image, the flame feature area is made more obvious and easy to identify; Step 3, by transforming the RGB three primary color channels of the ROI area processed above into HSI channels, the flame texture features are deepened, and the preprocessing process of the flame image is completed; Step 4: Divide the data set into the ROI area of ​​the HSI channel frame by frame, and put it into the Resnet101 convolutional neural network for training to obtain the weight file of converter flame detection; Step 5: Deploy the parameters and weights of the Resnet101 network, process the furnace mouth flame in the converter smelting process in real time, input it into the network for detection, and return the label value of the smelting period.

2. The converter endpoint determination method based on deep learning according to claim 1, characterized in that: In step 2, the flame recognition region of interest (ROI) of each frame image is adjusted and selected through the OpenCV open source image library and according to the flame change law of different converter smelting periods.

3. The converter endpoint determination method based on deep learning according to claim 2, characterized in that: Step 3 includes: extracting flame features from the original image frame, firstly by selecting the ROI area, and then completing the conversion of the RGB channel of the flame to the HSI channel according to the following calculation formulas (1)-(4); In the above formula, H represents the phase angle of color, which is the most important attribute of color and determines the essence of color; S represents the depth of color, the higher the saturation, the darker the color; I represents the brightness of the color, R represents the red channel; G represents the green channel; B represents the blue channel, and a multi-scale expression based on the texture and color characteristics of the converter flame is constructed.

4. The converter endpoint determination method based on deep learning according to claim 3, characterized in that: In step 4, the Pytorch deep learning framework is used to build the Resnet101 residual convolutional neural network. The number of residual layers is 101, and the label files are "early", "mid", and "late" respectively. The tensor dimensions of the input in the training process are unified as [3,244,244].

5. The converter endpoint determination method based on deep learning according to claim 3 or 4, characterized in that: A converter endpoint determination system is used, the system comprising: The acquisition module is used to obtain a continuous video stream of the flame state during the converter steelmaking process; The processing module performs image processing on the continuous video stream, including the selection of the flame region of interest and the conversion of the RGB channel of the ROI region to the HSI channel; The calculation module inputs the input picture frame into the network model after preprocessing, performs forward calculation through the weight file, and returns the index value of the preset label; The judgment module searches for the label and the corresponding confidence level through the preset status label index value to complete the judgment of the flame status at the smelting moment.

Images