Automatic slagging-off method and system for molten iron pretreatment of steelmaking desulfurization station

Abstract

The invention provides an automatic slagging-off method and system for molten iron pretreatment of a steelmaking desulfurization station, and relates to the technical field of ferrous metallurgy, and the method comprises the following steps: capturing image information of a slagging-off area in real time by using image acquisition equipment, and constructing an accurate molten iron slag block visual recognition model in combination with a traditional machine vision technology and a deep learning technology; and then the optimal path on which the slag rake can advance before slagging-off each time is calculated in real time through a path planning technology, so that automatic slagging-off operation is automatically, efficiently and accurately assisted, and the defects of manual slagging-off are effectively overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of iron and steel metallurgy. Specifically, it relates to an automatic slag skimming system for hot metal pretreatment in a steelmaking desulfurization station based on machine vision and path optimization algorithms, and particularly to an automatic slag skimming method and system for hot metal pretreatment in a steelmaking desulfurization station. Background Art

[0002] Currently, the slag skimming operation for hot metal pretreatment in a steelmaking desulfurization station is carried out manually. The operator observes and determines the slag skimming area with more slag blocks in the hot metal through the naked eye, and uses the handle to control the movement of the slag skimming machine arm (connected with the slag rake). The slag skimming machine arm is immersed into the molten iron surface of the ladle, and when the slag rake retracts, the iron slag is taken out from the slag outlet of the hot metal ladle. However, the on-site slag skimming operation environment has a high temperature, and a large amount of dust is generated during the slag skimming process. There are many potential safety hazards in the manual slag skimming method. The perception, decision-making, and control of the slag skimming operation are all completed manually, and the automation and intelligence levels of the slag skimming machine are low, which urgently need to be improved.

[0003] The existing invention patent with the publication number CN105353654A discloses an iron water slag skimming detection and control system based on image processing. This system monitors the slag skimming process in real time and judges the slag skimming grade online according to the slag skimming standard database. The invention patent with the publication number CN108986098A divides the real-time collected image into a first region of interest and a second region of interest in the disclosed technology, corresponding to processing two scenarios with and without a slag skimming shovel. The invention patent with the publication number CN112017145A discloses an efficient automatic slag skimming method for hot metal pretreatment. After analyzing the region and area of the hot metal slag in the image, the slag skimming head is automatically controlled to skim the slag along the optimal path. At the same time, a blowing slag driving device is also enabled to gather the scattered hot metal slag together for efficient slag skimming.

[0004] In the above-mentioned existing technologies, although all introduce the use of image recognition methods to recognize the initial pictures to obtain the images of the ladle mouth area of the hot metal, these methods are all based on traditional machine vision methods. In the case of different slag skimming environments in desulfurization stations and uncertain camera acquisition positions, it is extremely difficult to ensure accurate segmentation of the elliptical area of the ladle mouth at one time. Since the area to be detected of the iron slag cannot be accurately recognized, a series of subsequent steps such as recognizing hot metal and iron slag, calculating the optimal path of the slag rake, and assisting the slag skimming operator to skim the slag will not meet the initial requirements. Summary of the Invention

[0005] Aiming at the deficiencies in the prior art, the present invention provides an automatic slag skimming method and system for hot metal pretreatment in a steelmaking desulfurization station.

[0006] An automatic slag skimming method and system for hot metal pretreatment in a steelmaking desulfurization station according to the present invention are as follows:

[0007] In a first aspect, an automatic slag skimming method for hot metal pretreatment in a steelmaking desulfurization station is provided. The method includes:

[0008] Model training step: Training an iron ladle mouth area segmentation model;

[0009] Image acquisition step: In the hot metal pretreatment area of the steelmaking desulfurization station, real-time video image data of the hot metal pretreatment area of the desulfurization station is acquired;

[0010] Image processing step: Preprocessing the acquired video image data, inputting the processed picture into the iron ladle mouth area segmentation model to obtain the final elliptical area of the iron ladle mouth, and setting the elliptical area of the iron ladle mouth as the ROI for the subsequent slag block recognition step;

[0011] Slag block recognition step: Inputting the preprocessed image of the ROI area into a slag block area and area recognition unit mainly based on machine vision algorithms, and outputting the centroid coordinates of the area enriched with slag blocks on the current hot metal surface and the area size information of the slag blocks in the enriched area;

[0012] Path planning step: Planning the slag rake dropping point and the optimal path to be traveled through the output result of the previous step;

[0013] Control instruction step: Generating on-site control data for assisting automatic slag skimming according to the recognized current hot metal slag surface area and the optimal path of the slag rake traveling, and transmitting it to the control center server of the desulfurization station.

[0014] Preferably, the model training step includes:

[0015] Step S1.1: Collecting on-site pictures of different hot metal slag skimming areas in desulfurization stations, and screening these pictures to eliminate pictures with blurred shooting or inaccurate alignment with the slag skimming area;

[0016] Step S1.2: Labeling the screened picture materials using the LabelMe labeling tool; according to the area size of the iron ladle mouth in the whole picture, using different numbers of labeling amounts to segment and label the iron ladle mouth, and finally obtaining a labeling file containing the picture name and the coordinate information of the iron ladle mouth mask area;

[0017] Step S1.3: Constructing a convolutional neural network model for iron ladle mouth area recognition: Determining the input layer size of the convolutional neural network model, scaling the image to a fixed size according to the preprocessed image size, and setting the dimension of the input layer as width, height, and number of channels;

[0018] Design multiple convolutional layers. For the convolutional operation, use convolutional kernels of different sizes and numbers to perform convolution on the input feature map, and design the convolutional kernels as 3×3. After the convolutional operation, introduce the non-linear activation function ReLU and perform batch normalization on the output of the convolutional layer. After the convolutional layer, add a pooling layer for max pooling to downsample the feature map, reduce the size of the feature map, and retain the feature information at the same time. Design multiple convolutional-pooling modules, and introduce skip connections in each module to fuse low-level features and high-level features to improve the segmentation accuracy.

[0019] Use a transposed convolutional layer for upsampling to restore the feature map to the size of the original image. Finally, introduce a convolutional layer to map the upsampled feature map to the number of output channels corresponding to the number of classes in the segmentation task, where the elliptical area of the taphole is used as the foreground and other areas are used as the background. Through the softmax activation function, convert the output of the convolutional layer into the probability distribution of each pixel belonging to each class, so as to realize the pixel-level segmentation of the input image and obtain the final segmentation model of the elliptical area of the taphole.

[0020] Preferably, the image acquisition step includes: installing one or more industrial network cameras in the hot metal pretreatment area of the desulfurization station. The industrial network cameras are connected to the server of the desulfurization station control center through a wired or wireless network to transmit video image data in real time. The video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps.

[0021] Preferably, the image processing step includes:

[0022] Step S3.1: For the noise in the taphole hot metal slag surface image, use the median filtering algorithm to remove it. Let the input image be f(x,y), and the size of the filtering window be (2N + 1)×(2N + 1). Then the median filtering output g(x,y) is: g(x,y) = median f(x - i, y - j), i,j ∈ [-N,N]; where, median represents the median operation, N represents the size parameter of the filtering window, and i,j represent the pixel offsets within the filtering window.

[0023] Step S3.2: Select histogram equalization or the Retinex algorithm for image enhancement. Let the gray distribution of the input image be p(r k ), then the gray value s k after histogram equalization is: where, L is the number of gray levels, MN is the total number of image pixels, and n j represents the gray value r jThe number of pixels; The Retinex algorithm realizes image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x,y) = logS(x,y) - logpS(x,y)*G(x,y,σ); where R(x,y) is the reflection component, S(x,y) is the original image, and G(x,y,σ) is a Gaussian filter with a standard deviation of σ.

[0024] Step S3.3: Input the preprocessed picture into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, which will be set as the ROI of the subsequent slag block recognition unit; Assume the size of the original image is M×N, the upper left coordinate of the ROI area is (x0,y0), and the area of the area is m×n. Then the ROI extraction operation is expressed as: ROI = f(x0,y0).

[0025] Preferably, the slag block recognition steps include:

[0026] Step S4.1: Slag block area recognition, identifying the molten iron slag block area through image segmentation technology; Using binarization operation to convert the image into a form that only contains the target slag block and the background, where the target slag block appears as an area with a different gray value from the molten iron. Apply morphological operations to remove noise and fill the holes in the slag block, thereby defining the boundary of the slag block; Use connected region analysis to label and identify individual slag block areas, and each connected region represents an independent slag block.

[0027] Step S4.2: Slag block area recognition, calculating the number of pixels of each marked connected region as the area of the slag block; For each connected region, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average value of all pixel coordinates:

[0028]

[0029] where A i is the area of connected region i, C x and C y are the centroid coordinates of this area respectively; According to the actual application requirements, set an area threshold to identify the slag blocks in the enrichment area; The slag blocks with an area larger than this threshold are considered as the enrichment area, and finally output the area A i of each slag block in the enrichment area and the corresponding centroid point coordinates (C x , C y ).

[0030] Preferably, the path planning steps include:

[0031] Step S5.1: Determine the global connection and feasible paths: Using the coordinates of the centroid point, starting from the ladle slag tapping opening, based on the condition of non-backtracking in one direction, establish a global connection from the nodes with lower Y-axis to the nodes with higher Y-axis; for any two nodes i and j, if Y i <Y j , then establish a one-way connection from i to j between them, and construct the distance matrix D of the feasible paths between the nodes, where D ij represents the straight-line distance from node i to node j:

[0032]

[0033] where, X i and Y i are the abscissa and ordinate of node i respectively;

[0034] Step S5.2: Calculate the optimal path: Traverse the feasible paths from the slag tapping opening to each node, record the cumulative distance and cumulative slag surface area of each route, for each path, judge whether the cumulative slag surface area of the nodes it passes through is greater than or equal to the slag rake area, that is, ∑ k S k ≥θ, where θ is the slag rake area, to determine the optimal path;

[0035] If the cumulative slag surface area is greater than or equal to the slag rake area, the optimal path of the slag rake travel is expressed as:

[0036]

[0037] If the cumulative slag surface area is less than the slag rake area, the optimal path of the slag rake travel is expressed as:

[0038]

[0039] where, k is the number of all nodes included in the slag raking path, and S k is the cumulative amount of iron slag in the slag rake after traveling k nodes;

[0040] Step S5.3: Output the slag rake travel order: According to the result of the optimal path, reverse the arrangement order of the nodes to meet the travel direction of the slag rake starting from the slag tapping opening, and output the order of the nodes that the slag rake will pass through in turn, providing a navigation path for the slag rake control system.

[0041] Preferably, the control instruction steps include:

[0042] Step S6.1: Make a judgment based on the slag block area obtained in the slag block recognition step. If the proportion of the slag block area in the area of the whole ladle of hot metal is less than 15%, no slag skimming operation is performed and False is directly returned; if the proportion of the slag block area in the area of the whole ladle of hot metal exceeds 15%, calculate the centroid point of the slag block enrichment area and the optimal path of the slag rake.

[0043] Step S6.2: Pack the information including the flag of whether to skim slag (i.e., True or False), the value of the proportion of iron slag in the area of the whole ladle of hot metal, and the coordinate values of the traveling path of the slag rake into a structured data packet according to the set data format. The data packet includes header information: timestamp, data length, and check code to ensure the integrity and reliability of the data.

[0044] Step S6.3: Use the TCP / IP protocol to transmit the packed result data to the server of the desulfurization station control center in real time through industrial Ethernet or other reliable network protocols. Encryption and data compression measures are adopted during the transmission process to ensure the security and efficiency of the data.

[0045] Step S6.4: When the desulfurization station server receives the slag skimming data packet, parse the data format and store the data in a relational database or a time-series database for subsequent query, analysis, and application.

[0046] In the second aspect, an automatic slag skimming system for hot metal pretreatment in a steelmaking desulfurization station is provided. The system includes:

[0047] Model training module: Train the ladle mouth area segmentation model.

[0048] Image acquisition module: In the hot metal pretreatment area of the steelmaking desulfurization station, collect video image data of the hot metal pretreatment area of the desulfurization station in real time.

[0049] Image processing module: Preprocess the collected video image data, input the processed picture into the ladle mouth area segmentation model to obtain the final elliptical area of the ladle mouth, and set the elliptical area of the ladle mouth as the ROI for the subsequent slag block recognition module.

[0050] Slag block recognition module: Input the preprocessed image in the ROI area into the slag block area and area recognition unit mainly based on machine vision algorithms, and output the centroid point coordinates of the area where the slag blocks are enriched on the current hot metal surface and the area size information of the slag blocks in the enrichment area.

[0051] Path planning module: Plan the dropping point of the slag rake and the optimal path required to travel through the output result of the previous module.

[0052] Control Instruction Module: Generate on-site control data for assisting automatic slag skimming based on the recognized current molten iron slag surface area and the optimal path of the slag rake, and transmit it to the desulfurization station control center server.

[0053] Preferably, the model training module includes:

[0054] Module M1.1: Collect on-site pictures of the molten iron slag skimming areas of different desulfurization stations, and screen these pictures to eliminate pictures with blurred shooting or inaccurate alignment with the slag skimming area.

[0055] Module M1.2: Label the screened picture materials using the LabelMe annotation tool; according to the area size of the ladle mouth in the whole picture, use different numbers of annotation points to segment and label the ladle mouth, and finally obtain an annotation file containing the picture name and the coordinate information of the ladle mouth mask area.

[0056] Module M1.3: Construct a convolutional neural network model for ladle mouth area recognition: Determine the input layer size of the convolutional neural network model, scale the image to a fixed size according to the preprocessed image size, and set the dimensions of the input layer to width, height, and number of channels.

[0057] Design multiple convolutional layers. The convolution operation uses convolutional kernels of different sizes and numbers to perform convolution on the input feature map, and the convolutional kernel is designed as 3×3; after the convolution operation, introduce the non-linear activation function ReLU and perform batch normalization on the output of the convolutional layer; after the convolutional layer, add a pooling layer for max pooling to downsample the feature map, reduce the size of the feature map, and at the same time retain the feature information; design multiple convolutional-pooling modules, and introduce skip connections in each module to fuse low-level features and high-level features to improve the segmentation accuracy.

[0058] Use a deconvolution layer for upsampling to restore the feature map to the size of the original image; finally, introduce a convolutional layer to map the upsampled feature map to the number of output channels of the segmentation task category, where the elliptical area of the ladle mouth is used as the foreground and other areas are used as the background; convert the output of the convolutional layer into the probability distribution of each pixel belonging to each category through the softmax activation function, so as to achieve pixel-level segmentation of the input image and obtain the final segmentation model of the ladle mouth elliptical area.

[0059] Preferably, the image acquisition module includes: Install one or more industrial network cameras in the molten iron pretreatment area of the desulfurization station. The industrial network cameras are connected to the server of the desulfurization station control center through a wired or wireless network to transmit video image data in real time. The video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps.

[0060] The image processing module includes:

[0061] Module M3.1: For the noise in the image of the slag surface of the torpedo ladle containing molten iron, the median filtering algorithm is used for removal. Let the input image be f(x, y), and the size of the filtering window be (2N + 1)×(2N + 1). Then the median filtering output g(x, y) is: g(x, y) = median f(x - i, y - j), i, j ∈ [-N, N]; where, median represents the operation of taking the median value, N represents the size parameter of the filtering window, and i, j represent the pixel offsets within the filtering window;

[0062] Module M3.2: Select histogram equalization or the Retinex algorithm for image enhancement. Let the gray - level distribution of the input image be p(r k ), then the gray - level value s k after histogram equalization is: where, L is the number of gray - level steps, MN is the total number of image pixels, and n j represents the number of pixels with gray - level value r j ; The Retinex algorithm realizes image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x, y) = logS(x, y) - log[S(x, y)*G(x, y, σ)]; where, R(x, y) is the reflection component, S(x, y) is the original image, and G(x, y, σ) is a Gaussian filter with standard deviation σ;

[0063] Module M3.3: Input the pre - processed picture into the above - mentioned torpedo - ladle orifice region segmentation model to obtain the final elliptical region of the torpedo - ladle orifice, and this region will be set as the ROI of the subsequent slag - block recognition unit. Let the size of the original image be M×N, the upper - left coordinate of the ROI region be (x0, y0), and the size of the region be m×n. Then the ROI extraction operation is expressed as: ROI = f(x0, y0);

[0064] The slag - block recognition module includes:

[0065] Module M4.1: Slag - block region recognition. Identify the molten - iron slag - block region through image segmentation technology. Use binary operation to convert the image into a form that only contains the target slag - block and the background, where the target slag - block appears as a region with a different gray - level value from that of the molten iron. Apply morphological operations to remove noise and fill the holes within the slag - block, thereby accurately defining the boundary of the slag - block. Use connected - component analysis to label and identify individual slag - block regions, and each connected component represents an independent slag - block;

[0066] Module M4.2: Identification of the slag block area. Calculate the number of pixels in each labeled connected region as the area of the slag block. For each connected region, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average of all pixel coordinates:

[0067]

[0068] where A i is the area of the connected region i, and C x and C y are the centroid coordinates of this region respectively. According to the actual application requirements, set an area threshold to identify the slag blocks in the enrichment region. The slag blocks with an area larger than this threshold are considered as the enrichment region. Finally, output the area A i of each slag block in the enrichment region and the corresponding centroid point coordinates (C x , C y );

[0069] The path planning module includes:

[0070] Module M5.1: Determine the global connection and feasible path: Using the centroid point coordinates, starting from the slag discharge opening of the ladle, based on the condition of non-backtracking in one direction, establish a global connection from the lower node of the Y-axis to the higher node of the Y-axis. For any two nodes i and j, if Y i < Y j , then establish a one-way connection from i to j between them, and construct the distance matrix D of the feasible path between nodes, where D ij represents the straight-line distance from node i to node j:

[0071]

[0072] where X i and Y i are the horizontal and vertical coordinates of node i respectively;

[0073] Module M5.2: Calculate the optimal path: Traverse the feasible paths from the slag discharge opening to each node, record the cumulative distance and cumulative slag surface area of each route. For each path, judge whether the cumulative slag surface area of the nodes it passes through is greater than or equal to the area of the slag rake, that is, ∑ k S k ≥ θ, where θ is the area of the slag rake, to determine the optimal path;

[0074] If the cumulative slag surface area is greater than or equal to the area of the slag rake, the optimal path of the slag rake movement is expressed as:

[0075]

[0076] If the cumulative slag surface area is less than the slag rake area, the optimal path of the slag rake movement is expressed as:

[0077]

[0078] where k is the total number of all nodes included in the slag scraping path, and S k is the amount of iron slag accumulated in the slag rake after the slag rake moves through k nodes;

[0079] Module M5.3: Output the movement sequence of the slag rake: According to the result of the optimal path, reverse the arrangement order of the nodes to meet the movement direction of the slag rake starting from the slag discharge port, and output the sequence of nodes that the slag rake will pass through in turn, providing a navigation path for the slag rake control system;

[0080] The control instruction module includes:

[0081] Module M6.1: Judge according to the slag block area obtained in the slag block recognition module. If the proportion of the slag block area in the whole ladle of hot metal area is less than 15%, no slag scraping operation is performed and False is directly returned; if the proportion of the slag block area in the whole ladle of hot metal area exceeds 15%, calculate the centroid of the slag block enrichment area and the optimal path of the slag rake;

[0082] Module M6.2: Pack the information including the flag of whether to scrape slag (i.e., True or False), the value of the proportion of iron slag in the whole ladle of hot metal area, and the coordinate value of the slag rake movement path into a structured data packet according to the set data format. The data packet includes header information: timestamp, data length, check code, to ensure the integrity and reliability of the data;

[0083] Module M6.3: Use the TCP / IP protocol to transmit the packed result data to the server of the desulfurization station control center in real time through the industrial Ethernet or other reliable network protocols. Encryption and data compression measures are adopted during the transmission process to ensure the security and efficiency of the data;

[0084] Module M6.4: When the desulfurization station server receives the slag scraping data packet, parse the data format and store the data in a relational database or a time series database for subsequent query, analysis and application.

[0085] Compared with the prior art, the present invention has the following beneficial effects:

[0086] The present invention solves the problems of low automation and intelligence, lagging response, low efficiency, etc. in the traditional operation where the operator controls the arm of the slag scraping machine (connected with the slag rake) through the handle on the console to take out the iron slag immersed in the hot metal surface of the ladle from the slag discharge port of the hot metal ladle, significantly improves the automation level and accuracy of hot metal desulfurization and slag scraping, and thus effectively prevents the loss caused by the hot metal being taken out of the ladle due to human operation errors during the slag scraping process.

[0087] Other beneficial effects of the present invention will be described in the specific implementation manners by introducing specific technical features and technical solutions. Those skilled in the art should be able to understand the beneficial technical effects brought by the technical features and technical solutions through the introduction of these technical features and technical solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] By reading the following detailed description of non-limiting embodiments with reference to the accompanying drawings, other features, objects, and advantages of the present invention will become more apparent:

[0089] Figure 1 is the system flow chart of the present invention;

[0090] Figure 2 is the schematic diagram after preprocessing the original image collected in the ladle area;

[0091] Figure 3 is the schematic diagram of the ROI of the elliptical area at the ladle mouth. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0092] The present invention will be described in detail below with reference to specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that those of ordinary skill in the art can make several changes and improvements without departing from the concept of the present invention. These all belong to the protection scope of the present invention.

[0093] An embodiment of the present invention provides an automatic slag skimming method for hot metal pretreatment in a steelmaking desulfurization station. Referring to Figure 1 as shown, the method specifically includes:

[0094] Model training step: Train the ladle mouth area segmentation model. In order to accurately identify the elliptical area of the ladle mouth in different desulfurization station scenarios and provide good basic conditions for subsequent steps such as iron slag area and slag block area identification and slag rake path planning, this step trains the ladle mouth area segmentation model.

[0095] This step specifically includes:

[0096] Step S1.1: Use the cameras in the hot metal slag skimming areas of different desulfurization stations to collect a certain number of on-site pictures in each scenario, and manually screen these collected pictures to eliminate pictures with blurred shooting or inaccurate alignment with the slag skimming area. Finally, a total of 120 pictures in 3 scenarios are obtained.

[0097] Step S1.2: Label the 120 collected image materials using the LabelMe annotation tool. According to the proportion of the ladle mouth area in the whole image, different numbers of points are used for annotation to segment and label the ladle mouth. Finally, an annotation file containing the image name and the coordinate information of the ladle mouth mask area is obtained.

[0098] Step S1.3: Build a convolutional neural network model for ladle mouth area recognition: First, determine the input layer size of the convolutional neural network model. According to the size of the preprocessed image, scale the image to a fixed size, such as 256×256, to ensure that the model input sizes are consistent.

[0099] Considering the number of color channels of the image, the dimension of the input layer is set to (width, height, number of channels). Design multiple convolutional layers for extracting local features of the image. Each convolutional layer includes a convolution operation, an activation function, and batch normalization. The convolution operation uses convolution kernels of different sizes and numbers to perform convolution on the input feature map, and the convolution kernel is designed as 3×3. After the convolution operation, introduce the non-linear activation function ReLU to enhance the non-linear expression ability of the network. Perform batch normalization on the output of the convolutional layer to accelerate the network training process and improve the generalization ability of the model. After the convolutional layer, add a pooling layer for max pooling to downsample the feature map, reduce the size of the feature map, and retain important feature information at the same time.

[0100] To enhance the model's perception ability of features at different scales, design multiple convolution-pooling modules, and introduce skip connections in each module to fuse low-level features with high-level features and improve the segmentation accuracy. Next, use a transposed convolution (deconvolution) layer for upsampling to restore the feature map to the size of the original image. Finally, introduce a convolutional layer to map the upsampled feature map to the number of output channels corresponding to the number of categories in the segmentation task (the elliptical area of the ladle mouth as the foreground and other areas as the background). Convert the output of the convolutional layer into the probability distribution of each pixel belonging to each category through the softmax activation function, so as to achieve pixel-level segmentation of the input image and obtain the final segmentation model of the elliptical area of the ladle mouth.

[0101] Image acquisition step: In the hot metal pretreatment area of the steelmaking desulfurization station, video image data of the hot metal pretreatment area of the desulfurization station is collected in real time. This step specifically includes: Install industrial-grade network cameras in the appropriate area of the hot metal pretreatment in the steelmaking desulfurization station, and the image acquisition unit collects the video image data of the hot metal pretreatment area of the desulfurization station in real time. Install one or more industrial-grade network cameras in the hot metal pretreatment area of the desulfurization station. The cameras have characteristics such as dust-proof, moisture-proof, and high-temperature resistance, and can work stably in harsh environments. The cameras are connected to the server in the desulfurization station control center through wired or wireless networks to transmit high-definition video image data in real time. The video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps to ensure that the image quality meets the requirements for hot metal slag surface recognition.

[0102] Image processing step: Preprocess the collected video image data, including operations such as image grayscale processing, image denoising, and image enhancement. The preprocessed image should be able to fully display the characteristics of the ladle mouth area, providing high-quality input for subsequent recognition of the hot metal slag area, as shown in Figure 2 shown. Input the processed picture into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, and set this ladle mouth elliptical area as the ROI for the subsequent slag block recognition step, as shown in Figure 3 shown.

[0103] This step specifically includes:

[0104] Step S3.1: For the noise points in the ladle hot metal slag surface image, use the median filtering algorithm to remove them. The basic principle of the median filtering algorithm is to replace the gray value of a pixel point with the median of the gray values in the neighborhood of the pixel point, which has a good effect on removing impulse noise and salt-and-pepper noise. Let the input image be f(x,y), and the size of the filtering window be (2N + 1)×(2N + 1), then the median filtering output g(x,y) is: g(x,y) = median f(x - i, y - j), i, j ∈ [-N, N]; where, median represents the operation of taking the median, N represents the size parameter of the filtering window, and i, j represent the pixel offsets within the filtering window;

[0105] Step S3.2: In order to improve the contrast and detail information of the image, choose histogram equalization or the Retinex algorithm for image enhancement. Histogram equalization enhances the contrast of the image by stretching the gray distribution of the image so that it is evenly distributed within the entire gray range. Let the gray distribution of the input image be p(r k ), then the gray value s k after histogram equalization is: where, L is the number of gray levels, MN is the total number of image pixels, and n j represents the gray value of r jThe number of pixels; The Retinex algorithm realizes image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x, y) = logS(x, y) - log[S(x, y) * G(x, y, σ)]; where R(x, y) is the reflection component, S(x, y) is the original image, and G(x, y, σ) is a Gaussian filter with a standard deviation of σ.

[0106] Step S3.3: Input the preprocessed picture into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, which will be set as the ROI of the subsequent slag block recognition unit to reduce the amount of data for subsequent image processing and exclude the interference of other areas, thereby effectively improving the accuracy and speed of recognition. Let the size of the original image be M×N, the upper left coordinate of the ROI area be (x0, y0), and the area size be m×n. Then the ROI extraction operation is expressed as: ROI = f(x0, y0).

[0107] Slag block recognition steps: Input the preprocessed image (only containing the ladle mouth elliptical ROI area) into the slag block area and area recognition unit mainly based on traditional machine vision algorithms. This unit mainly uses traditional image analysis and machine vision algorithms to output information such as the centroid coordinates of the area enriched with slag blocks on the current molten iron surface and the area size of the slag blocks in the enriched area. This step specifically includes:

[0108] Step S4.1: Slag block area recognition. First, identify the molten iron slag block area through image segmentation technology. Image segmentation can be achieved through methods such as threshold segmentation, edge detection, or connected region analysis. Use binary operation to convert the image into a form that only contains the target slag blocks and the background, where the target slag blocks appear as areas with different gray values from the molten iron. Apply morphological operations such as opening operation and closing operation to remove noise and fill the holes in the slag blocks, thereby clearly defining the boundaries of the slag blocks. Use connected component analysis (CCA) to label and identify individual slag block areas. Each connected region represents an independent slag block, [labeled image = CCA(binary image)].

[0109] Step S4.2: Slag block area recognition (recognition of slag block enriched area). First, calculate the number of pixels of each labeled connected region as the area of the slag block. For each connected region, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average value of all pixel coordinates:

[0110]

[0111]

[0112] Among them, A i is the area of the connected region i, and C x and C y are the centroid coordinates of this region respectively; according to the actual application requirements, an area threshold is set to identify the slag blocks in the enrichment region; the slag blocks with an area larger than this threshold are considered as the enrichment region, and finally the area A i of each slag block in the enrichment region and the corresponding centroid point coordinates (C x , C y ) are output.

[0113] Path planning step: Based on the output result of the previous step, plan the dropping point of the slag rake and the optimal path to be traveled. This step specifically includes:

[0114] Step S5.1: Determine the global connection and the feasible path: Using the centroid point coordinates, with the tapping spout of the ladle as the starting point, based on the condition of one-way non-backtracking, establish a global connection from the lower node of the Y-axis to the higher node of the Y-axis; for any two nodes i and j, if Y i < Y j , then establish a one-way connection from i to j between them, and construct the distance matrix D of the feasible paths between the nodes, where D ij represents the straight-line distance from node i to node j:

[0115]

[0116] Among them, X i and Y i are the horizontal and vertical coordinates of node i respectively;

[0117] Step S5.2: Calculate the optimal path: Traverse the feasible paths from the tapping spout to each node, record the cumulative distance and the cumulative slag surface area of each route, and for each path, judge whether the cumulative slag surface area of the nodes it passes through is greater than or equal to the area of the slag rake, that is, ∑ k S k ≥ θ, where θ is the area of the slag rake, to determine the optimal path;

[0118] If the cumulative slag surface area is greater than or equal to the area of the slag rake, the optimal path for the slag rake to travel is expressed as:

[0119]

[0120] If the cumulative slag surface area is less than the area of the slag rake, the optimal path for the slag rake to travel is expressed as:

[0121]

[0122] Among them, k is the number of all nodes included in the slag raking path, and S kis the cumulative amount of iron slag in the slag rake after it travels k nodes;

[0123] Step S5.3: Output the slag rake travel sequence: According to the result of the optimal path, reverse the arrangement order of the nodes to meet the travel direction of the slag rake starting from the slag discharge port, and output the sequence of nodes that the slag rake will pass through in turn, providing a navigation path for the slag rake control system.

[0124] Control instruction step: Generate on-site control data for auxiliary automatic slag skimming based on the recognized current molten iron slag surface area and the result of the optimal path coordinate position of the slag rake travel, and transmit it to the desulfurization station control center server. This step specifically includes:

[0125] Step S6.1: Make a judgment based on the slag block area obtained in the slag block recognition step. If the proportion of the slag block area in the whole ladle of molten iron area is less than 15% (this value can be adjusted according to the actual situation), no slag skimming operation is performed and False is directly returned; if the proportion of the slag block area in the whole ladle of molten iron area exceeds 15%, calculate the centroid point of the slag block enrichment area and the optimal path of the slag rake;

[0126] Step S6.2: Pack the information including the slag skimming flag (True or False), the value of the iron slag proportion in the whole ladle of molten iron area, and the slag rake travel path coordinate value into a structured data packet according to the set data format. The data packet includes header information: timestamp, data length, check code, etc., to ensure the integrity and reliability of the data;

[0127] Step S6.3: Use the TCP / IP protocol to transmit the packed result data to the server of the desulfurization station control center through the industrial Ethernet or other reliable network protocols in real time. Encryption and data compression measures are adopted during the transmission process to ensure the security and efficiency of the data;

[0128] Step S6.4: When the desulfurization station server receives the slag skimming data packet, parse the data format and store the data in a relational database or a time series database for subsequent query, analysis and application.

[0129] The present invention also provides an automatic slag skimming system for hot metal pretreatment in a steelmaking desulfurization station. The automatic slag skimming system for hot metal pretreatment in the steelmaking desulfurization station can be implemented by executing the process steps of the automatic slag skimming method for hot metal pretreatment in the steelmaking desulfurization station. That is, those skilled in the art can understand the automatic slag skimming method for hot metal pretreatment in the steelmaking desulfurization station as a preferred implementation manner of the automatic slag skimming system for hot metal pretreatment in the steelmaking desulfurization station. This system specifically includes:

[0130] Model training module: Train the ladle mouth area segmentation model. In order to accurately identify the elliptical area of the ladle mouth in different desulfurization station scenarios and provide good basic conditions for subsequent modules such as iron slag area and slag block area identification and slag rake path planning, this module trains the ladle mouth area segmentation model.

[0131] This module specifically includes:

[0132] Module M1.1: Use the cameras in the hot metal slag skimming areas of different desulfurization stations to collect a certain number of on-site pictures in each scenario, and manually screen these collected pictures to eliminate pictures with blurred shooting or inaccurate alignment with the slag skimming area. Finally, a total of 120 pictures in 3 scenarios are obtained.

[0133] Module M1.2: Use the LabelMe annotation tool to annotate the 120 collected picture materials. According to the area size of the ladle mouth in the whole picture, different numbers of annotation points are used to segment and annotate the ladle mouth, and finally an annotation file containing the picture name and the coordinate information of the ladle mouth mask area is obtained.

[0134] Module M1.3: Build a convolutional neural network model for ladle mouth area recognition: First, determine the input layer size of the convolutional neural network model. According to the size of the preprocessed image, scale the image to a fixed size, such as 256×256, to ensure that the model input size is consistent.

[0135] Considering the number of color channels of the image, the dimension of the input layer is set to (width, height, number of channels). Design multiple convolutional layers to extract local features of the image. Each convolutional layer includes a convolution operation, an activation function, and batch normalization. The convolution operation uses convolution kernels of different sizes and numbers to perform convolution on the input feature map, and the convolution kernel is designed as 3×3. After the convolution operation, introduce the non-linear activation function ReLU to enhance the non-linear expression ability of the network. Perform batch normalization on the output of the convolutional layer to accelerate the network training process and improve the generalization ability of the model. After the convolutional layer, add a pooling layer for max pooling to downsample the feature map, reduce the size of the feature map, and retain important feature information at the same time.

[0136] To enhance the model's perception ability of features at different scales, multiple convolution-pooling modules are designed, and skip connections are introduced in each module to fuse low-level features with high-level features, improving the segmentation accuracy. Next, a transposed convolution layer is used for upsampling to restore the feature map to the size of the original image. Finally, a convolution layer is introduced to map the upsampled feature map to the number of output channels corresponding to the number of classes in the segmentation task (the elliptical area of the tuyere as the foreground and other areas as the background). The output of the convolution layer is transformed into the probability distribution of each pixel belonging to each class through the softmax activation function, thereby achieving pixel-level segmentation of the input image and obtaining the final segmentation model for the elliptical area of the tuyere.

[0137] Image acquisition module: In the molten iron pretreatment area of the steelmaking desulfurization station, video image data of the molten iron pretreatment area of the desulfurization station is collected in real time. Specifically, this module includes: installing industrial-grade network cameras in appropriate areas of the molten iron pretreatment in the steelmaking desulfurization station, and the image acquisition unit collects video image data of the molten iron pretreatment area of the desulfurization station in real time. One or more industrial-grade network cameras are installed in the molten iron pretreatment area of the desulfurization station. The cameras have characteristics such as dust-proof, moisture-proof, and high-temperature resistance, and can work stably in harsh environments. The cameras are connected to the server in the desulfurization station control center through wired or wireless networks to transmit high-definition video image data in real time. The video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps to ensure that the image quality meets the requirements of molten iron slag surface recognition.

[0138] Image processing module: Preprocess the collected video image data, including operations such as video frame splitting, image denoising, image enhancement, and image segmentation. The preprocessed image should be able to fully display the characteristics of the tuyere area, providing high-quality input for subsequent recognition of the molten iron slag area. The processed picture is input into the tuyere area segmentation model to obtain the final elliptical area of the tuyere, and this elliptical area of the tuyere is set as the ROI of the subsequent slag block recognition module.

[0139] Specifically, this module includes:

[0140] Module M3.1: For the noise points in the molten iron slag surface image of the tuyere, a median filtering algorithm is used for removal. The basic principle of the median filtering algorithm is to replace the gray value of a pixel point with the median of the gray values in the neighborhood of the pixel point, which has a good effect on removing impulse noise and salt-and-pepper noise. Let the input image be f(x,y), and the size of the filtering window be (2N + 1)×(2N + 1), then the median filtering output g(x,y) is: g(x,y) = median f(x - i, y - j), i, j ∈ [-N, N]; where, median represents the operation of taking the median, N represents the size parameter of the filtering window, and i, j represent the pixel offsets within the filtering window;

[0141] Module M3.2: To improve the contrast and detail information of the image, histogram equalization or the Retinex algorithm is selected for image enhancement. Histogram equalization enhances the image contrast by stretching the gray-scale distribution of the image to make it evenly distributed across the entire gray-scale range. Let the gray-scale distribution of the input image be p(r k ), then the gray-scale value s k after histogram equalization is: where L is the number of gray levels, MN is the total number of image pixels, and n j represents the number of pixels with gray-scale value r j ; The Retinex algorithm realizes image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x,y) = logS(x,y) - log[S(x,y)*G(x,y,σ)]; where R(x,y) is the reflection component, S(x,y) is the original image, and G(x,y,σ) is a Gaussian filter with standard deviation σ;

[0142] Module M3.3: Input the preprocessed picture into the ladle mouth area segmentation model to obtain the final elliptical ladle mouth area, which will be set as the ROI of the subsequent slag block recognition unit to reduce the data volume of subsequent image processing, exclude the interference of other areas, and thus effectively improve the recognition accuracy and speed. Let the size of the original image be M×N, the upper-left coordinate of the ROI area be (x0,y0), and the area size be m×n. Then the ROI extraction operation is expressed as: ROI = f(x0,y0).

[0143] Slag block recognition module: Input the preprocessed image (only containing the elliptical ladle mouth ROI area) into the slag block area and area recognition unit mainly based on traditional machine vision algorithms. This unit mainly uses traditional image analysis and machine vision algorithms to output information such as the centroid coordinates of the area enriched with slag blocks on the current molten iron surface and the area size of the slag blocks in the enriched area. This module specifically includes:

[0144] Module M4.1: Slag block area recognition. First, identify the molten iron slag block area through image segmentation technology. Image segmentation can be achieved through methods such as threshold segmentation, edge detection, or connected component analysis. Use binary operation to convert the image into a form that only contains the target slag blocks and the background, where the target slag blocks appear as areas with different gray-scale values from the molten iron. Apply morphological operations such as opening operation and closing operation to remove noise and fill the holes in the slag blocks, so as to clearly define the boundaries of the slag blocks. Adopt connected component analysis (Connected Component Analysis, CCA) to label and identify individual slag block areas. Each connected component represents an independent slag block, [labeled image = CCA(binary image)].

[0145] Module M4.2: Identification of slag block area (identification of slag block enrichment area). First, calculate the number of pixels of each labeled connected region, which is used as the area of the slag block. For each connected region, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average value of all pixel coordinates:

[0146]

[0147]

[0148] where A i is the area of connected region i, and C x and C y are the centroid coordinates of this region respectively; according to the actual application requirements, set an area threshold to identify the slag blocks in the enrichment area; the slag blocks with an area greater than this threshold are considered as the enrichment area, and finally output the area A i of each slag block in the enrichment area and the corresponding centroid point coordinates (C x , C y ).

[0149] Path planning module: Based on the output results of the previous module, plan the dropping point of the slag rake and the optimal path to be traveled. This module specifically includes:

[0150] Module M5.1: Determine the global connection and feasible path: Using the centroid point coordinates, starting from the slag discharge outlet of the ladle, based on the condition of non-backtracking in one direction, establish a global connection from the lower node of the Y-axis to the higher node of the Y-axis; for any two nodes i and j, if Y i < Y j , then establish a one-way connection from i to j between them, and construct the distance matrix D of the feasible path between nodes, where D ij represents the straight-line distance from node i to node j:

[0151]

[0152] where X i and Y i are the horizontal and vertical coordinates of node i respectively;

[0153] Module M5.2: Calculate the optimal path: Traverse the feasible paths from the slag discharge outlet to each node, record the cumulative distance and cumulative slag surface area of each route, and for each path, judge whether the cumulative slag surface area of the nodes it passes through is greater than or equal to the area of the slag rake, that is, ∑ k S k ≥ θ, where θ is the area of the slag rake, to determine the optimal path;

[0154] If the cumulative slag surface area is greater than or equal to the slag rake area, the optimal path for the slag rake to travel is expressed as:

[0155]

[0156] If the cumulative slag surface area is less than the slag rake area, the optimal path for the slag rake to travel is expressed as:

[0157]

[0158] where k is the total number of nodes included in the slag scraping path, and S k is the amount of iron slag accumulated in the slag rake after the slag rake travels k nodes;

[0159] Module M5.3: Output the traveling order of the slag rake: According to the result of the optimal path, reverse the arrangement order of the nodes to meet the traveling direction of the slag rake starting from the slag outlet, and output the order of the nodes that the slag rake will pass through in sequence, providing a navigation path for the slag rake control system.

[0160] Control instruction module: Generate on-site control data for assisting automatic slag scraping according to the recognized current molten iron slag surface area and the result of the optimal path coordinate position of the slag rake traveling, and transmit it to the desulfurization station control center server. This module specifically includes:

[0161] Module M6.1: Make a judgment according to the slag block area obtained in the slag block recognition module. If the proportion of the slag block area in the area of the whole ladle of molten iron is less than 15% (this value can be adjusted according to the actual situation), no slag scraping operation is performed and False is directly returned; if the proportion of the slag block area in the area of the whole ladle of molten iron exceeds 15%, calculate the centroid point of the slag block enrichment area and the optimal path of the slag rake;

[0162] Module M6.2: Pack the information including the slag scraping flag (True or False), the value of the iron slag proportion in the area of the whole ladle of molten iron, and the coordinate values of the slag rake traveling path into a structured data packet according to the set data format. The data packet includes header information: timestamp, data length, check code, etc., to ensure the integrity and reliability of the data;

[0163] Module M6.3: Use the TCP / IP protocol to transmit the packed result data to the server of the desulfurization station control center through the industrial Ethernet or other reliable network protocols in real time. Encryption and data compression measures are adopted during the transmission process to ensure the security and efficiency of the data;

[0164] Module M6.4: When the desulfurization station server receives the slag scraping data packet, parse the data format and store the data in a relational database or a time series database for subsequent query, analysis and application.

[0165] An embodiment of the present invention provides an automatic slag skimming method and system for hot metal pretreatment in a steelmaking desulfurization station. In the segmentation algorithm for the elliptical area at the ladle mouth (i.e., the area where hot metal and slag are distributed), a deep learning algorithm based on a neural network architecture is adopted. By collecting a certain number of images under multiple different desulfurization station scenarios and accurately annotating these images, the elliptical mask area of the ladle mouth is segmented. Finally, combined with the annotation materials and the original images, a model capable of accurately segmenting the elliptical area of the ladle mouth under different desulfurization station scenarios is trained, laying a foundation for subsequent steps such as identifying the slag area and its area, and planning the slag rake path, so as to realize automatic, efficient, and accurate auxiliary automatic slag skimming operations and effectively make up for the deficiencies of manual slag skimming.

[0166] Those skilled in the art know that in addition to implementing the system and its various devices, modules, and units provided by the present invention in the form of pure computer-readable program code, the method steps can be logically programmed to enable the system and its various devices, modules, and units provided by the present invention to be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers, etc. to achieve the same functions. Therefore, the system and its various devices, modules, and units provided by the present invention can be regarded as a hardware component, and the devices, modules, and units included therein for implementing various functions can also be regarded as the structures within the hardware component; the devices, modules, and units for implementing various functions can also be regarded as either software modules for implementing the method or structures within the hardware component.

[0167] The specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other arbitrarily.

Claims

1. An automatic slag removal method for molten iron pretreatment in a steelmaking desulfurization station, characterized in that: include: Model training steps: training the ladle mouth area segmentation model; Image acquisition steps: in the hot metal pretreatment area of ​​the steelmaking desulfurization station, real-time video image data of the hot metal pretreatment area of ​​the desulfurization station is collected; Image processing step: pre-processing the collected video image data, inputting the processed image into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, and setting the ladle mouth elliptical area as the ROI of the subsequent slag block recognition step; Slag block identification step: input the preprocessed image of the ROI area into the slag block area and area identification unit based on the machine vision algorithm, and output the coordinates of the center of gravity of the slag block enrichment area on the current molten iron surface and the area size information of the slag block in the enrichment area; Path planning step: Based on the output results of the previous step, plan the slag rake landing point and the optimal path required for travel; Control instruction steps: Based on the identified current molten iron slag surface area and the optimal path for the slag rake to move, generate on-site control data for auxiliary automatic slag removal and transmit it to the desulfurization station control center server.

2. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The model training step includes: Step S1.1: Collect on-site pictures of the molten iron slagging area at different desulfurization stations, and screen these pictures to remove pictures that are blurred or not accurately aligned with the slagging area; Step S1.2: Label the filtered image materials using the LabelMe annotation tool; segment and annotate the iron ladle opening using different annotation points according to the area of ​​the iron ladle opening in the entire image, and finally obtain an annotation file containing the image name and the coordinate information of the iron ladle opening mask area; Step S1.3: Construct a convolutional neural network model for iron ladle mouth area recognition: determine the input layer size of the convolutional neural network model, scale the image to a fixed size according to the preprocessed image size, and set the dimensions of the input layer to width, height, and number of channels; Design multiple convolutional layers. The convolution operation uses convolution kernels of different sizes and numbers to convolve the input feature map. The convolution kernel is designed to be 3×3. After the convolution operation, the nonlinear activation function ReLU is introduced to perform batch normalization on the output of the convolution layer. After the convolution layer, a pooling layer is added to perform maximum pooling to downsample the feature map and reduce the size of the feature map while retaining the feature information. Design multiple convolution-pooling modules and introduce skip connections in each module to fuse low-level features with high-level features to improve segmentation accuracy. The deconvolution layer is used for upsampling to restore the feature map to the size of the original image. Finally, a convolution layer is introduced to map the upsampled feature map to the output channel number of the number of categories of the segmentation task, where the iron clad mouth elliptical area is used as the foreground and other areas are used as the background. The output of the convolution layer is converted into the probability distribution of each pixel belonging to each category through the softmax activation function, thereby realizing pixel-level segmentation of the input image and obtaining the final iron clad mouth elliptical area segmentation model.

3. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The image acquisition step includes: installing one or more industrial-grade network cameras in the molten iron pretreatment area of ​​the desulfurization station, and the industrial-grade network cameras are connected to the server of the desulfurization station control center through a wired or wireless network to transmit video image data in real time. The video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps.

4. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The image processing step comprises: Step S3.1: Use median filtering algorithm to remove noise points in the slag surface image of the iron ladle; assuming that the input image is f(x,y), the filter window size is (2N+1)×(2N+1), then the median filter output g(x,y) is: g(x,y)=median f(xi,yj),i,j∈[-N,N]; where median represents the median operation, N represents the size parameter of the filter window, and i,j represents the pixel offset within the filter window; Step S3.2: Select histogram equalization or Retinex algorithm for image enhancement. Suppose the grayscale distribution of the input image is p(r k ), then the gray value s after histogram equalization k for: Among them, L is the gray level, MN is the total number of image pixels, and n j Indicates the gray value is r j The Retinex algorithm achieves image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x,y)=logS(x,y)-log[S(x,y)*G(x,y,σ)]; where R(x,y) is the reflection component, S(x,y) is the original image, and G(x,y,σ) is a Gaussian filter with a standard deviation of σ. Step S3.3: Input the preprocessed image into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, which will be set as the ROI of the subsequent slag block identification unit; assuming that the original image size is M×N, the coordinates of the upper left corner of the ROI area are (x0, y0), and the area size is m×n, then the ROI extraction operation is expressed as: ROI=f(x0, y0).

5. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The slag block identification step comprises: Step S4.1: Slag block area identification, using image segmentation technology to identify the molten iron slag block area; using a binarization operation to convert the image into a form containing only the target slag block and the background, where the target slag block appears as an area with a different grayscale value from the molten iron, and applying morphological operations to remove noise and fill holes in the slag block, thereby accurately defining the boundaries of the slag block; using connected region analysis to mark and identify separate slag block areas, each connected region represents an independent slag block; Step S4.2: Identify the area of ​​the slag block. Calculate the number of pixels in each marked connected area as the area of ​​the slag block. For each connected area, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average value of all pixel coordinates: Among them, A i is the area of ​​connected region i, C x and C y are the centroid coordinates of the region respectively; according to the actual application requirements, an area threshold is set to identify the slag blocks in the enriched area; the slag blocks with an area greater than the threshold are considered to be enriched areas, and finally the area A of the slag blocks in each enriched area is output i And the corresponding center of gravity coordinates (C x ,C y ).

6. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The path planning step includes: Step S5.1: Determine global connection and feasible path: Using the coordinates of the centroid point, starting from the ladle slag outlet, and based on the one-way non-backtracking condition, establish a global connection from the lower Y-axis node to the higher Y-axis node; for any two nodes i and j, if Y i <Y j , then a one-way connection from i to j is established between them, and the distance matrix D of the feasible paths between nodes is constructed, where D ij Represents the straight-line distance from node i to node j: Among them, X i and Y i are the horizontal and vertical coordinates of node i respectively; Step S5.2: Calculate the optimal path: traverse the feasible paths from the slag outlet to each node, record the cumulative distance and cumulative slag surface area of ​​each route, and for each path, determine whether the cumulative slag surface area of ​​the nodes it passes through is greater than or equal to the slag rake area, that is, ∑ k S k ≥θ, where θ is the slag rake area, to determine the optimal path; If the accumulated slag surface area is greater than or equal to the slag rake area, the optimal path for the slag rake to move is expressed as: If the accumulated slag surface area is smaller than the slag rake area, the optimal path for the slag rake to move is expressed as: Among them, k is the number of nodes contained in the scrapping path, S k is the amount of iron slag accumulated in the slag rake after the slag rake has traveled k nodes; Step S5.3: Output the moving order of the slag rake: According to the result of the optimal path, the arrangement order of the nodes is inverted to meet the moving direction of the slag rake starting from the slag outlet, and the order of the nodes that the slag rake will pass through in sequence is output to provide a navigation path for the slag rake control system.

7. The automatic slagging method for molten iron pretreatment in a steelmaking desulfurization station according to claim 1, characterized in that: The control instruction step comprises: Step S6.1: judging according to the slag area obtained in the slag identification step, if the slag area accounts for less than 15% of the whole ladle of molten iron area, no slag removal operation is performed and False is directly returned; if the slag area accounts for more than 15% of the whole ladle of molten iron area, the centroid of the slag enrichment area and the optimal path of the slag rake are calculated; Step S6.2: According to the set data format, the information including the flag of whether to scrape slag, i.e. True or False, the ratio of iron slag to the area of ​​the whole package of molten iron, and the coordinate value of the slag rake's travel path are packaged into a structured data packet, and the data packet includes header information: timestamp, data length, and checksum to ensure the integrity and reliability of the data; Step S6.3: Using TCP / IP protocol, the packaged result data is transmitted in real time to the server of the desulfurization station control center through industrial Ethernet or other reliable network protocols. Encryption and data compression measures are used in the transmission process to ensure data security and efficiency; Step S6.4: After receiving the slagging data packet, the desulfurization station server parses the data format and stores the data in a relational database or a time series database to facilitate subsequent query, analysis and application.

8. An automatic slag removal system for hot metal pretreatment in a steelmaking desulfurization station, characterized in that: include: Model training module: training the segmentation model of the iron ladle mouth area; Image acquisition module: In the hot metal pretreatment area of ​​the steelmaking desulfurization station, real-time video image data of the hot metal pretreatment area of ​​the desulfurization station is collected; Image processing module: pre-processing the collected video image data, inputting the processed image into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, and setting the ladle mouth elliptical area as the ROI of the subsequent slag block recognition module; Slag block recognition module: input the pre-processed image of the ROI area into the slag block area and area recognition unit based on machine vision algorithm, and output the coordinates of the center of gravity of the slag block enrichment area on the current molten iron surface and the area size information of the slag block in the enriched area; Path planning module: plans the slag rake landing point and the optimal path required based on the output results of the previous module; Control instruction module: Based on the identified current molten iron slag surface area and the optimal path for the slag rake to move, it generates on-site control data for auxiliary automatic slag removal and transmits it to the desulfurization station control center server.

9. The automatic slag removal system for hot metal pretreatment in a steelmaking desulfurization station according to claim 8, characterized in that: The model training module includes: Module M1.1: Collect on-site pictures of the hot metal slagging area at different desulfurization stations, and screen these pictures to remove those that are blurred or not accurately aligned with the slagging area; Module M1.2: Use the LabelMe annotation tool to annotate the filtered image materials; according to the area size of the iron ladle mouth in the entire image, use different number of annotation points to segment and annotate the iron ladle mouth, and finally obtain an annotation file containing the image name and the coordinate information of the iron ladle mouth mask area; Module M1.3: Construct a convolutional neural network model for iron ladle mouth area recognition: Determine the input layer size of the convolutional neural network model, scale the image to a fixed size based on the preprocessed image size, and set the dimensions of the input layer to width, height, and number of channels; Design multiple convolutional layers. The convolution operation uses convolution kernels of different sizes and numbers to convolve the input feature map. The convolution kernel is designed to be 3×3. After the convolution operation, the nonlinear activation function ReLU is introduced to perform batch normalization on the output of the convolution layer. After the convolution layer, a pooling layer is added to perform maximum pooling to downsample the feature map and reduce the size of the feature map while retaining the feature information. Design multiple convolution-pooling modules and introduce skip connections in each module to fuse low-level features with high-level features to improve segmentation accuracy. The deconvolution layer is used for upsampling to restore the feature map to the size of the original image. Finally, a convolution layer is introduced to map the upsampled feature map to the output channel number of the number of categories of the segmentation task, where the iron clad mouth elliptical area is used as the foreground and other areas are used as the background. The output of the convolution layer is converted into the probability distribution of each pixel belonging to each category through the softmax activation function, thereby realizing pixel-level segmentation of the input image and obtaining the final iron clad mouth elliptical area segmentation model.

10. The automatic slag removal system for hot metal pretreatment in a steelmaking desulfurization station according to claim 8, characterized in that: The image acquisition module includes: installing one or more industrial-grade network cameras in the hot metal pretreatment area of ​​the desulfurization station, the industrial-grade network cameras are connected to the server of the desulfurization station control center through a wired or wireless network, and real-time transmission of video image data, the video acquisition resolution is not less than 1920×1080, and the frame rate is not less than 25fps; The image processing module comprises: Module M3.1: The median filter algorithm is used to remove the noise points in the slag surface image of the iron ladle. Assuming the input image is f(x,y), the filter window size is (2N+1)×(2N+1), then the median filter output g(x,y) is: g(x,y)=median f(xi,yj),i,j∈[-N,N]; where median means the median operation, N means the size parameter of the filter window, and i,j means the pixel offset within the filter window. Module M3.2: Select histogram equalization or Retinex algorithm for image enhancement. Suppose the grayscale distribution of the input image is p(r k ), then the gray value s after histogram equalization k for: Among them, L is the gray level, MN is the total number of image pixels, and n j Indicates the gray value is r j The Retinex algorithm achieves image dynamic range compression and contrast enhancement by estimating the illumination component of the image and removing it from the original image. The basic steps are: logR(x,y)=logS(x,y)-log[S(x,y)*G(x,y,σ)]; where R(x,y) is the reflection component, S(x,y) is the original image, and G(x,y,σ) is a Gaussian filter with a standard deviation of σ. Module M3.3: Input the preprocessed image into the ladle mouth area segmentation model to obtain the final ladle mouth elliptical area, which will be set as the ROI of the subsequent slag block recognition unit; assuming that the original image size is M×N, the coordinates of the upper left corner of the ROI area are (x0, y0), and the area size is m×n, then the ROI extraction operation is expressed as: ROI=f(x0, y0); The slag block identification module comprises: Module M4.1: Slag region identification, using image segmentation technology to identify the molten iron slag region; using binarization operation to convert the image into a form containing only the target slag and background, where the target slag is represented by a region with a different gray value from the molten iron, and applying morphological operations to remove noise and fill holes in the slag, thereby accurately defining the boundaries of the slag; using connected region analysis to mark and identify individual slag regions, each of which represents an independent slag; Module M4.2: Slag block area identification, calculate the number of pixels in each marked connected area, and use it as the area of ​​the slag block; for each connected area, calculate the average value of its pixel coordinates to determine the centroid coordinates of the slag block. The centroid coordinates are calculated by taking the average value of all pixel coordinates: Among them, A i is the area of ​​connected region i, C x and C y are the centroid coordinates of the region respectively; according to the actual application requirements, an area threshold is set to identify the slag blocks in the enriched area; the slag blocks with an area greater than the threshold are considered to be enriched areas, and finally the area A of the slag blocks in each enriched area is output i And the corresponding center of gravity coordinates (C x ,C y ); The path planning module includes: Module M5.1: Determine global connection and feasible path: Using the coordinates of the center of gravity, starting from the ladle slag outlet, and based on the one-way non-backtracking condition, establish a global connection from the lower Y-axis node to the higher Y-axis node; for any two nodes i and j, if Y i <Y j , then a one-way connection from i to j is established between them, and the distance matrix D of the feasible paths between nodes is constructed, where D ij Represents the straight-line distance from node i to node j: Among them, X i and Y i are the horizontal and vertical coordinates of node i respectively; Module M5.2: Calculate the optimal path: traverse the feasible paths from the slag outlet to each node, record the cumulative distance and cumulative slag surface area of ​​each route, and for each path, determine whether the cumulative slag surface area of ​​the nodes it passes through is greater than or equal to the slag rake area, that is, ∑ k S k ≥θ, where θ is the slag rake area, to determine the optimal path; If the accumulated slag surface area is greater than or equal to the slag rake area, the optimal path for the slag rake to move is expressed as: If the accumulated slag surface area is smaller than the slag rake area, the optimal path for the slag rake to move is expressed as: Among them, k is the number of nodes contained in the scrapping path, S k is the amount of iron slag accumulated in the slag rake after the slag rake has traveled k nodes; Module M5.3: Output the moving order of the slag rake: According to the result of the optimal path, the arrangement order of the nodes is inverted to meet the moving direction of the slag rake starting from the slag outlet, and the order of the nodes that the slag rake will pass through is output to provide a navigation path for the slag rake control system; The control instruction module comprises: Module M6.1: According to the slag area obtained in the slag identification module, if the slag area accounts for less than 15% of the entire ladle of molten iron, no slag removal operation is performed and False is directly returned; if the slag area accounts for more than 15% of the entire ladle of molten iron, the centroid of the slag enrichment area and the optimal path of the slag rake are calculated; Module M6.2: According to the set data format, the information including whether to remove slag, True or False, the proportion of slag in the whole package of molten iron, and the coordinate value of the slag rake's travel path are packaged into a structured data packet. The data packet includes header information: timestamp, data length, and checksum to ensure the integrity and reliability of the data; Module M6.3: Using TCP / IP protocol, the packaged result data is transmitted to the server of the desulfurization station control center in real time through industrial Ethernet or other reliable network protocols. Encryption and data compression measures are adopted in the transmission process to ensure data security and efficiency; Module M6.4: After the desulfurization station server receives the slagging data packet, it parses the data format and stores the data in a relational database or a time series database to facilitate subsequent query, analysis and application.

Images