Steel billet positioning control method, device, electronic equipment and computer-readable storage medium

Abstract

The present application discloses a method, device, electronic device and computer-readable storage medium for positioning and controlling a steel billet before it enters the furnace. The method according to the embodiments of the present application includes: obtaining an original image of a target steel billet before it enters the furnace; performing image processing on the original image to obtain a processed image; inputting the processed image into a trained network model to obtain the position information of the target steel billet; and based on the position information of the target steel billet and preset specified position information, adjusting the motion state of the roller table for carrying the target steel billet, thereby controlling the target steel billet to move to the specified position. This solution utilizes machine vision and deep learning technologies to achieve automatic positioning of steel billets of different sizes before they enter the furnace and achieve the effect of precise positioning.

Description

Technical Field

[0001] This application relates to the technical field of automatic control of heating furnaces, and particularly to a method, device, electronic device, and computer-readable storage medium for positioning control before a steel billet enters a furnace. Background Art

[0002] A steel billet is a product obtained by casting molten steel produced in a steelmaking furnace. Steel billets can be mainly divided into two types from the manufacturing process: ingot-cast billets and continuous-cast billets. Steel billets can be mainly divided into two types from the appearance: slab billets: the ratio of the width to the height of the cross-section is relatively large, mainly used for rolling plates; square billets: the width and height of the cross-section are equal or not much different, mainly used for rolling sections and wire rods. Steel billets can also be called steel products. After being processed, they can be used as mechanical parts, forgings, and various steel products processed, and wire rods are the functions of steel billets. Further, a steel billet refers to a semi-finished product used for producing steel products and generally cannot be directly used.

[0003] At present, before a steel billet enters a furnace, it needs to be accurately positioned so that the steel billets are positioned and arranged according to the designed charging pattern to meet the requirements of its heating process, thereby ensuring the outlet temperature of the steel billets. Since the temperature in front of the heating furnace is very high, it is difficult to install corresponding detection devices. Therefore, the positioning work of the steel billets in front of the furnace needs to be achieved through manual judgment, and manual adjustment is required during this process. In this way, it is difficult to ensure the accuracy of the steel billet positioning, and it is easy to affect the steel charging rhythm, thereby resulting in a relatively low overall production efficiency. Summary of the Invention

[0004] The purpose of this application is that embodiments of this application provide a method, device, electronic device, and computer-readable storage medium for positioning control before a steel billet enters a furnace, aiming to utilize machine vision and deep learning technologies to achieve automatic positioning of steel billets of different sizes before entering the furnace and achieve the effect of accurate positioning.

[0005] According to one aspect of this application, an embodiment of this application provides a method for positioning control before a steel billet enters a furnace, which includes: acquiring an original image of a target steel billet before entering the furnace; performing image processing on the original image to obtain a processed image; inputting the processed image into a trained network model to obtain the position information of the target steel billet; and based on the position information of the target steel billet and preset specified position information, adjusting the motion state of the roller table for carrying the target steel billet, thereby controlling the target steel billet to move to the specified position.

[0006] Optionally, in some embodiments of this application, before the step of inputting the processed image into the trained network model, it includes: collecting a sample original image of a sample steel billet before entering the furnace; performing calibration processing on the sample original image to obtain a calibrated sample image and the actual label of the sample steel billet; and based on the calibrated sample image and the actual label, training a preset network model to obtain a trained network model.

[0007] Optionally, in some embodiments of the present application, the step of calibrating the original sample image to obtain the calibrated sample image and the actual label of the sample billet includes: performing image preprocessing on the calibrated sample image.

[0008] Optionally, in some embodiments of the present application, the step of training the preset network model based on the calibrated sample image and the actual label to obtain the trained network model includes: initializing the model parameters of the preset network model; generating training samples according to the calibrated sample image; training the preset network model based on the training samples to obtain a training result, and extracting the predicted label of the sample billet based on the training result; comparing the predicted label and the actual label of the sample billet to obtain a comparison result; adjusting the model parameters of the preset network model based on the comparison result until the preset network model converges to obtain the trained network model.

[0009] Optionally, in some embodiments of the present application, the preset network model includes an input node, an intermediate node, and an output node, where the input node is used to receive training samples to obtain input tensor data; the output node is used to obtain output tensor data and output a training result accordingly; the intermediate node includes multiple levels of downsampling nodes and multiple levels of upsampling nodes; each level of downsampling node is connected to the next level of downsampling node through a corresponding residual downsampling module, and each level of downsampling node is used to obtain the intermediate tensor data of the current level of downsampling node according to the input tensor data and the intermediate tensor data of the previous level of downsampling node; each level of upsampling node is connected to the next level of upsampling node through a corresponding residual upsampling module, and each level of upsampling node is used to obtain the intermediate tensor data of the current level of upsampling node according to the intermediate tensor data of the last level of downsampling node and the intermediate tensor data of the previous level of upsampling node; each level of downsampling node is cross-connected to the corresponding upsampling node.

[0010] Optionally, in some embodiments of the present application, the number of the residual downsampling modules is used to represent the number of downsampling levels, and the value range of the number of downsampling levels is m≤N≤n, where N is the number of downsampling levels, and m and n are the first threshold and the second threshold of the number of downsampling levels. The first threshold is determined according to the receptive field of the preset network model, and the second threshold is determined according to the receptive field of the preset network model and the size of the sample image.

[0011] Optionally, in some embodiments of the present application, an additional node is further provided between the last level of upsampling node and the output node, and the additional tensor data of the additional node is obtained based on the intermediate tensor data of the last level of upsampling node and performing at least one of convolution operation, normalization operation, and activation function operation.

[0012] Optionally, in some embodiments of the present application, the output tensor data of the output node is obtained based on additional tensor data and by performing an adaptive average pooling operation.

[0013] Optionally, in some embodiments of the present application, the residual downsampling module includes a first branch and a second branch; the first branch is used to perform at least one of a convolution operation, a normalization operation, and an activation function operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path, and the second branch is used to obtain a second branch result for performing a superimposing operation and an activation function operation with the first branch result.

[0014] Optionally, in some embodiments of the present application, the residual upsampling module includes a first branch and a second branch; the first branch is used to perform at least one of a convolution operation, a normalization operation, and an activation function operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path, and the second branch is used to obtain a second branch result for performing a superimposing operation, an activation function operation, and an upsampling operation with the first branch result.

[0015] Optionally, in some embodiments of the present application, the shortcut path is used to perform a convolution operation and a normalization operation on the intermediate tensor data.

[0016] According to another aspect of the present application, an embodiment of the present application provides a billet positioning control device before furnace charging, the device includes: an image acquisition module, configured to acquire an original image of a target billet before furnace charging; an image processing module, configured to perform image processing on the original image to obtain a processed image; an information acquisition module, configured to input the processed image into a trained network model to obtain the position information of the target billet; and a billet moving module, configured to adjust the motion state of a roller table for carrying the target billet based on the position information of the target billet and preset specified position information, thereby controlling the target billet to move to the specified position.

[0017] According to another aspect of the present application, an embodiment of the present application provides an electronic device, the electronic device includes a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the billet positioning control method according to any embodiment of the present application are implemented.

[0018] According to still another aspect of the present application, an embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the steps of the billet positioning control method according to any embodiment of the present application are implemented.

[0019] The embodiments of the present application provide a method, device, electronic device, and computer-readable storage medium for pre-furnace positioning control of steel billets. By using machine vision and a deep learning network model, automatic positioning of steel billets with different sizes and types before entering the furnace can be achieved, thereby achieving the effect of precise positioning. Moreover, by reasonably designing the deep learning network model (or simply referred to as the network model, the same hereinafter), the problems of gradient explosion and network degradation can be better solved to accelerate the network learning process, and further the generalization problem of the deep network can be improved. Further, the deep learning network model adopts 5-level downsampling to increase the receptive field of the network model, and can also meet the requirements of positioning accuracy and real-time calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The following will clearly show the technical solutions and other beneficial effects of the present application by describing the specific embodiments of the present application in detail with reference to the accompanying drawings.

[0021] Figure 1 FIG. is a schematic diagram of a scenario of a method for pre-furnace positioning control of steel billets provided by an embodiment of the present application.

[0022] Figure 2 FIG. is a schematic flowchart of a method for pre-furnace positioning control of steel billets provided by an embodiment of the present application.

[0023] Figure 3 For implementing the Figure 2 FIG. is a schematic diagram of an industrial camera and a light source used when implementing the method for pre-furnace positioning control of steel billets as shown.

[0024] Figure 4 For Figure 2 FIG. is a schematic flowchart of the previous steps of step S300 as shown.

[0025] Figure 5 For Figure 4 FIG. is a schematic flowchart of the sub-steps of step S530 as shown.

[0026] Figure 6 For implementing the Figure 2 FIG. is a schematic structural diagram of the deep learning network model used when implementing the method for pre-furnace positioning control of steel billets as shown.

[0027] Figure 7 For Figure 6 FIG. is a schematic structural diagram of the residual downsampling module as shown.

[0028] Figure 8 For Figure 6 FIG. is a schematic structural diagram of the residual upsampling module as shown.

[0029] Figure 9 For Figure 7 or Figure 8Schematic structural diagram of the shortcut path shown.

[0030] Figure 10 Block diagram of the structure of the billet positioning control device before furnace charging provided by an embodiment of the present application.

[0031] Figure 11 Block diagram of the structure of the billet positioning control device before furnace charging provided by another embodiment of the present application.

[0032] Figure 12 For Figure 11 Block diagram of the structure of the sample image training module shown.

[0033] Figure 13 Schematic structural diagram of an electronic device provided by an embodiment of the present application. Detailed implementation manners

[0034] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the protection scope of the present application.

[0035] The embodiments of the present application provide a billet positioning control method, device, computer-readable storage medium and electronic device before furnace charging. Among them, the billet positioning control device before furnace charging can be integrated in the electronic device, and the electronic device can be a terminal or a server and other devices. The terminal can be a terminal device such as a smart phone, a tablet computer, a notebook computer, a personal computer (Personal Computer, abbreviated as PC), etc. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery network (Content Delivery Network, abbreviated as CDN), and big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected through wired or wireless communication methods, and the present application does not limit this here.

[0036] Please refer to Figure 1 , Figure 1 Schematic diagram of a scenario of a billet positioning control method provided by an embodiment of the present application.

[0037] Specifically, the above-mentioned electronic device is a server, which can acquire the original image of the target billet before entering the furnace; perform image processing on the original image to obtain a processed image; input the processed image into a trained network model to obtain the position information of the target billet; based on the position information of the target billet and the preset specified position information, adjust the motion state of the roller table for carrying the target billet, and thereby control the target billet to move to the specified position. In this way, by using machine vision and a deep learning network model, automatic positioning of billets with different sizes before entering the furnace can be achieved, and an accurate positioning effect can be achieved.

[0038] Correspondingly, an embodiment of the present application provides a method for controlling the positioning of a billet before entering the furnace, which includes: acquiring the original image of the target billet before entering the furnace; performing image processing on the original image to obtain a processed image; inputting the processed image into a trained network model to obtain the position information of the target billet; based on the position information of the target billet and the preset specified position information, adjust the motion state of the roller table for carrying the target billet, and thereby control the target billet to move to the specified position. By implementing the steps of the above method, precise positioning of billets with different sizes before entering the furnace can be achieved, thereby avoiding affecting the steel charging rhythm and improving production efficiency.

[0039] It should be noted that the method for controlling the positioning of a billet before entering the furnace provided by the embodiment of the present application involves machine vision and deep learning in the field of artificial intelligence.

[0040] The so-called artificial intelligence (AI for short) is to use a digital computer or a machine controlled by a digital computer to simulate, extend, and expand human intelligence, and is a theory, method, technology, and application system that can perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a comprehensive technology in computer science, which attempts to understand the essence of intelligence and produce a new intelligent machine that can react in a way similar to human intelligence. Artificial intelligence also studies the design principles and implementation methods of various intelligent machines, enabling the machines to have the functions of perception, reasoning, and decision-making. The artificial intelligence technology is an interdisciplinary subject, involving a wide range of fields, including both hardware-level technologies and software-level technologies. The basic artificial intelligence technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technology, operation / interaction systems, and mechatronics. The artificial intelligence software technologies mainly include several major directions such as machine vision technology, speech processing technology, natural language processing technology, and machine learning / deep learning, autonomous driving, and intelligent transportation.

[0041] Among them, machine vision technology is a technology that uses machines to replace human eyes for measurement and judgment. It converts the target to be captured into an image signal through machine vision products, transmits it to a dedicated image processing system, obtains the morphological information of the captured target, and converts it into a digital signal according to information such as pixel distribution, brightness, and color; the image processing system performs various operations on these signals to extract the features of the target, and then controls the actions of on-site equipment according to the discrimination results. By using machine vision technology, the flexibility and automation of production can be improved. Especially in some dangerous working environments that are not suitable for manual operation or occasions where manual vision is difficult to meet the requirements, machine vision is used to replace manual vision. At the same time, in the process of large-scale repetitive industrial production, machine vision technology can improve production efficiency and automation.

[0042] Among them, deep learning is generally a method of representing learning of data with an artificial neural network as the architecture. Deep learning is to learn the internal laws and representation levels of sample data, and the information obtained in these learning processes is very helpful for the interpretation of data such as text, images, and sounds. Its ultimate goal is to enable machines to have the ability of analysis and learning like humans, and be able to recognize text, image, and sound data. The advantage of deep learning is to use supervised, unsupervised, or semi-supervised feature learning and hierarchical feature extraction efficient algorithms to replace manual feature acquisition.

[0043] The following will be described in detail respectively. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments.

[0044] The embodiment of the present application provides a method for positioning and controlling a steel billet before entering the furnace. This method can be executed by a terminal or a server. The embodiment of the present application takes the method for positioning and controlling a steel billet before entering the furnace being executed by the server as an example for description.

[0045] Please refer to Figure 2 , Figure 2 , which is a schematic flowchart of a method for positioning and controlling a steel billet before entering the furnace provided by an embodiment of the present application. The specific process of this method for positioning and controlling a steel billet before entering the furnace can be as follows:

[0046] Step S100, obtain the original image of the target steel billet before entering the furnace.

[0047] In this step, the target steel billet before entering the furnace can be photographed by using an industrial camera to obtain the original image.

[0048] Since the temperature in front of the heating furnace is extremely high, if an ordinary camera is used to photograph the target billet before it enters the furnace for a long time, it may cause damage to the ordinary camera, thus affecting its normal operation. Therefore, in this embodiment, an industrial camera with good heat resistance and strong shooting performance is used to photograph the target billet and obtain relevant images of the target billet. In this way, it is not only not easily affected by the use environment, but also the photographed images can meet the requirements for image information extraction.

[0049] Refer to Figure 3 , Figure 3 It is a schematic diagram of the industrial camera and light source used when implementing the billet positioning control method before entering the furnace. During the process of photographing the target billet before it enters the furnace, it is necessary to preset the industrial camera (hereinafter referred to as the camera for short) and the light source in advance. The camera and the light source are set at a certain distance directly opposite the furnace door on the furnace inlet side, and the fields of view of the camera and the light source can cover the range where the target billet needs to be positioned, that is, the target area. In addition, when the billet production stops, a calibration grid is set on the route where the billet passes, and this calibration grid is used to indicate the position information of the billet and the position information of the preset specified points.

[0050] Step S200, perform image processing on the original image to obtain a processed image.

[0051] In this step, the image processing may include image grayscale conversion, but is not limited to image grayscale conversion, and may also include image sharpening, image noise reduction, etc. It should be noted that image grayscale conversion, image noise reduction, and image sharpening will be further described below.

[0052] Step S300, input the processed image into the trained network model to obtain the position information of the target billet.

[0053] In this step, the processed image is input into the trained network model, and the trained network model outputs the label of the target billet. This label includes the labels of both ends of the target billet (i.e., the start-end label and the end-end label). Then, operations for information extraction can be performed on the label to obtain the position information of the target billet. Specifically, the operation for information extraction can be to first obtain the start-end label and the end-end label of the target billet, and then project the start-end label and the end-end label onto the preset calibration grid, and the position information of the start end of the target billet and the position information of the end end of the target billet can be obtained, that is, the position information of the target billet. Further, according to the position information of the start end and the end end of the target billet, the length information of the target billet can be calculated.

[0054] Step S400, based on the position information of the target billet and the preset specified position information, adjust the movement state of the roller table for carrying the target billet, and accordingly control the target billet to move to the specified position.

[0055] Based on the position information of the target billet obtained in step S300 (including the position information of the starting end and the ending end of the target billet) and the preset specified position information, the position offset between the two positions can be obtained. After the server obtains the position offset, it can transmit the position offset to the on-site controller. Based on this, the on-site controller can control the motion state of the roller table for carrying the target billet, such as the moving speed of the roller table, moving forward at a speed of 0.1 m / s, so that the target billet continues to move. At the same time, the industrial camera is continuously turned on to take pictures of the target billet and the furnace door, and the original image of the target billet before entering the furnace obtained in real time is transmitted to the server. The server obtains the position information of the target billet in real time through the trained network model. When the target billet moves to the preset specified position (for example, the starting end of the target billet is aligned with the preset specified position point), the on-site controller controls the roller table to stop moving, thus realizing the precise positioning control before the billet enters the furnace, and also avoiding affecting the steel loading rhythm due to manual judgment of positioning, which can also improve production efficiency.

[0056] It should be noted that in the process of executing step 300 (that is, "input the processed image into the trained network model to obtain the position information of the target billet"), a network model is used. This network model is an artificial neural network architecture that can perform representation learning on data and output the learning results. In other words, this network model can also be called a deep learning network model. The specific structure of this network model will be further described below.

[0057] In order to improve the learning efficiency of this network model, before the step of "inputting the processed image into the trained network model", the following steps can be executed to train this network model and obtain the trained network model.

[0058] Refer to Figure 4 , Figure 4 For Figure 2 the schematic flow chart of the previous steps of step S300 shown. In this embodiment, before step S300, the method may include:

[0059] Step S510, collect the sample original image of the sample billet before entering the furnace.

[0060] To train the network model, it is necessary to collect samples. Specifically, after setting up the camera, light source, and other network devices, the actual sample images in the on-site production process can be collected using a preset sample collection program. In addition, to make the samples representative and diverse, when generating different steel grades, a large number of sample images generated throughout the day or for a longer time can be collected to meet the quantity requirements of the samples, so as to better train the network model. In addition, by selecting an industrial camera with a high-quality lens or using a hardware device with a high-performance configuration, high-quality sample images can be obtained to meet the quality requirements of the samples and further improve the training accuracy of the network model.

[0061] Step S520: Calibrate the original sample image to obtain the calibrated sample image and the actual label of the sample billet.

[0062] Optionally, step S520 further includes performing image preprocessing on the calibrated sample image.

[0063] Exemplarily, the preprocessing may include image grayscale conversion. Image grayscale conversion is to set the pixel points on the image to a sampled color. Grayscale images are usually displayed as grayscales from the darkest black to the brightest white. For example, the calibrated sample image (which is a color image) can be converted into a grayscale image through the following formula: Gray = R * 0.299 + G * 0.587 + B * 0.114, where Gray represents grayscale, R represents red pixels, G represents green pixels, and B represents blue pixels. It should be noted that the method of converting a color image into a grayscale image is not limited to this.

[0064] In this way, through the above-mentioned image grayscale conversion, the edge processing of the image can be effectively realized and the expressiveness of the image can be improved.

[0065] In other partial embodiments, the preprocessing may also include image denoising, which is also known as image noise reduction. Image denoising can adopt one of the methods of mean filter, adaptive Wiener filter, median filter, and morphological noise filter to reduce the noise in the image.

[0066] In other partial embodiments, the preprocessing may also include image sharpening. Exemplarily, the image sharpening is Laplacian operator image sharpening. After the image is grayscaled, the Laplacian operator image sharpening is performed on the grayscale image to highlight the obvious contour curves in the image and hide the relatively unobvious image boundaries. Laplacian is a second-order differential operator, so it emphasizes the sudden change in grayscale in the image and does not emphasize the slow-changing areas of the image. In this way, some gradually changing light gray edges will become the background color of the image contour. Specifically, the Laplacian sharpened image is related to the degree of mutation from the surrounding pixels of a certain pixel in the image to this pixel, that is, the basis of the Laplacian sharpened image is the degree of change of the image pixels. The first-order differential of a function describes where the function image changes, that is, grows or decreases; while the second-order differential describes the speed of image change, whether it is a sharp increase or decrease or a gentle increase or decrease. Based on this, it can be predicted that the transition degree of the pigments in the image can be found according to the second-order differential. For example, the transition from white to black is relatively sharp. Or rather, when the grayscale of the central pixel in the neighborhood is lower than the average grayscale of other pixels in its neighborhood, the grayscale of this central pixel should be further reduced, and when the grayscale of the central pixel in the neighborhood is higher than the average grayscale of other pixels in its neighborhood, the grayscale of this central pixel should be further increased, so as to achieve image sharpening.

[0067] It should be noted that it is necessary to confirm whether image preprocessing is required for the calibrated sample image according to the actual situation. If no image preprocessing is required, the calibrated sample image (color image) can be directly input into the preset network model. If image preprocessing is required, first convert the calibrated sample image (color image) into a grayscale image, and then input the grayscale image into the preset network model.

[0068] Continue to refer to Figure 4 , step S530, based on the calibrated sample image and the actual label, train the preset network model to obtain the trained network model.

[0069] Further refer to Figure 5 , Figure 5 For Figure 4 is a schematic flowchart of the sub-steps of step S530 shown. Step S530 further includes steps S531 to S535.

[0070] Step S531, initialize the model parameters of the preset network model.

[0071] Step S532, generate training samples according to the calibrated sample image.

[0072] Step S533, train the preset network model based on the training samples to obtain a training result, and extract the predicted label of the sample billet based on the training result.

[0073] In this embodiment, the preset network model is a deep learning network model. Refer to Figure 6 , Figure 6 which is a schematic structural diagram of the deep learning network model used when implementing the billet positioning control method before entering the furnace as shown in Figure 2 . The deep learning network model includes an input node X0, intermediate nodes (X1 to X13), and an output node X14. The input node X0 is used to receive training samples and obtain input tensor data; the output node X14 is used to obtain output tensor data and output training results accordingly; the intermediate nodes (X1 to X13) include multiple levels of downsampling nodes (X1 to X6) and multiple levels of upsampling nodes (X7 to X13); each level of downsampling node (X1 to X6) is connected to the next level of downsampling node through a corresponding residual downsampling module (C0 to C5). Each level of downsampling node (X1 to X6) is used to obtain the intermediate tensor data of the current level of downsampling node according to the input tensor data and the intermediate tensor data of the previous level of downsampling node; each level of upsampling node (X7 to X13) is connected to the next level of upsampling node through a corresponding residual upsampling module (UC1 to UC5). Each level of upsampling node (X7 to X13) is used to obtain the intermediate tensor data of the current level of upsampling node according to the intermediate tensor data of the last level of downsampling node and the intermediate tensor data of the previous level of upsampling node; each level of downsampling node (X1 to X6) is cross-connected to the corresponding upsampling node (X7 to X13). It should be noted that the cross-connection will be further described below.

[0074] In this network model, an additional node is also provided between the last level of upsampling node and the output node. The additional tensor data of the additional node is obtained based on the intermediate tensor data of the last level of upsampling node and performs at least one of convolution operation, normalization operation, and activation function operation.

[0075] Exemplarily, as shown in Figure 6 , the input node of the network model can be represented by X0, the output node by X14, and the intermediate nodes by X1 to X13. The input node X0 is used to receive training samples, which are grayscale images with width W * height H, and use the training samples as input tensor data. The output node X14 is used to generate and output training results after obtaining the output tensor data. The training results include a vector with length W. The intermediate nodes X1 to

[0076] Each of X11 obtains the intermediate tensor data generated in the network model. These intermediate tensor data can be represented by three dimensional parameters. The three dimensional parameters are width, height, and number of channels respectively. Exemplarily, the three dimensional parameters (width, height, number of channels) of the intermediate node X1 are set to (W, H, 4); the three dimensional parameters (width, height, number of channels) of the intermediate node X2 are set to (W / 2, H / 2, 8); the three dimensional parameters (width, height, number of channels) of the intermediate node X3 are set to (W / 4, H / 4, 16); the three dimensional parameters (width, height, number of channels) of the intermediate node X4 are set to (W / 8, H / 8, 32); the three dimensional parameters (width, height, number of channels) of the intermediate node X5 are set to (W / 16, H / 16, 64); the three dimensional parameters (width, height, number of channels) of the intermediate node X6 are set to (W / 32, H / 32, 64); the three dimensional parameters (width, height, number of channels) of the intermediate node X7 are set to (W / 16, H / 16, 128); the three dimensional parameters (width, height, number of channels) of the intermediate node X8 are set to (W / 8, H / 8, 64); the three dimensional parameters (width, height, number of channels) of the intermediate node X9 are set to (W / 4, H / 4, 32); the three dimensional parameters (width, height, number of channels) of the intermediate node X10 are set to (W / 2, H / 2, 16); the three dimensional parameters (width, height, number of channels) of the intermediate node X11 are set to (W, H, 8). In addition, the additional nodes can also be used as intermediate nodes. Therefore, the additional tensor data can also be represented by these three dimensional parameters (width, height, number of channels). Exemplarily, the three dimensional parameters (width, height, number of channels) of the additional node X12 are set to (W, H, 8); the three dimensional parameters (width, height, number of channels) of the additional node X13 are set to (W, H, 1).

[0077] Among them, intermediate nodes X1 to X6 are used to perform downsampling image processing on training samples, that is, subsampling or downsampling, and intermediate nodes X7 to X11 are used to perform upsampling image processing on training samples, that is, upsampling or image interpolation. Therefore, intermediate nodes X1 to X6 are correspondingly downsampling nodes X1 to X6; intermediate nodes X7 to X11 are correspondingly upsampling nodes X7 to X11. In this embodiment, when downsampling, the setting rule of these two dimension parameters (width, height) is to decrease exponentially, and at the same time, the setting rule of the remaining dimension parameter (number of channels) is to increase exponentially; when upsampling, the setting rule of these two dimension parameters (width, height) is to increase exponentially, and at the same time, the setting rule of the remaining dimension parameter (number of channels) is to decrease exponentially. Of course, in other partial embodiments, when downsampling or upsampling, the setting rules of the three dimension parameters (width, height, number of channels) can also be set to non-exponential. It should be noted that when downsampling, if the width and height are reduced, the number of channels needs to be increased to retain sufficient information. If the number of training samples is large enough and the computational load is not considered, the number of channels can be set very large. In actual operation, however, the training samples may not be sufficient, or the complexity of the network model needs to be simplified. Therefore, the number of channels cannot be set arbitrarily. When the information loss can be guaranteed to be small, the number of channels is set relatively small, which can not only ensure the generalization performance of the network model, but also reduce the inference time of the network model.

[0078] In this embodiment, each level of downsampling node is cross-connected with the corresponding upsampling node. As Figure 6 shown, downsampling node X1 is cross-connected with upsampling node X11; downsampling node X2 is cross-connected with upsampling node X10; downsampling node X3 is cross-connected with upsampling node X9; downsampling node X4 is cross-connected with upsampling node X8; downsampling node X5 is cross-connected with upsampling node X7; downsampling node X6 is connected to upsampling node X7 at the same layer. At the same time, each level of upsampling nodes X7 to X11 are connected in sequence. In other words, in this embodiment, the network model adopts bidirectional interpolation upsampling, which can make the quality of the enlarged image higher.

[0079] Furthermore, since the downsampling node and the upsampling node are cross-connected, the cross-connection splices the channel number dimension of the downsampling node and the channel number dimension of the upsampling node. For example, if the dimension parameter of the downsampling node is (W, H, C1) and the dimension parameter of the upsampling node is (W, H, C2), then the dimension parameter of the upsampling node after cross-connection is (W, H, (C1 + C2)).

[0080] In addition, it should be noted that since the output tensor data of the output node X14 is a one-dimensional vector (a vector of length W as described above), that is, the number of channels dimension in the output tensor data is 1, it is necessary to set the number of channels dimension of the intermediate node connected to the output node X14 to 1. When the last upsampling node in the multi-level upsampling node is connected to the output node, the number of channels dimension of the last upsampling node is set to 1, and the number of channels dimensions of the remaining upsampling nodes can be set to decrease or remain unchanged. When there is an additional node between the last upsampling node and the output node, and the last additional node in the additional node is connected to the output node, the number of channels dimension of the last additional node is set to 1. As Figure 6 shown, the number of channels dimension of the last additional node X13 connected to the output node X14 is set to 1.

[0081] Referring to Figure 6 and Figure 7 , Figure 7 is Figure 6 a schematic structural diagram of the residual downsampling module shown. As described above, a residual downsampling module is provided between each level of downsampling nodes. The residual downsampling module (C0 to C5) includes a first branch and a second branch; the first branch is used to perform at least one of the operations of convolution Conv operation, normalization BN operation, and activation function ReLU operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path Shortcut, and the second branch is used to obtain a second branch result to perform a superposition operation and an activation function operation with the first branch result. As Figure 7 and Figure 9As shown, in this embodiment, the residual downsampling module (C0 to C5) includes a first branch and a second branch; the first branch is used to sequentially perform a convolution Conv operation, a normalization BN operation, a ReLU activation function operation, a convolution Conv operation, and a normalization BN operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path Shortcut, which performs a convolution operation and a normalization operation on the intermediate tensor data to obtain a second branch result, and performs an addition operation and a ReLU activation function operation with the first branch result. Among them, when the first branch of the residual downsampling module (C0 to C5) performs the convolution Conv operation, the convolution operation uses a 3×3 convolution kernel, the stride is 1, and the padding is 1. For the second branch of the residual downsampling module (C0 to C5), that is, the shortcut path Shortcut, when performing the convolution Conv operation, the convolution operation uses a 1×1 convolution kernel, and the stride is determined according to whether downsampling is performed. For example, the convolution stride of the shortcut path in the residual downsampling module C0 between the input node X0 and the first-level downsampling node X1 is 1, indicating no downsampling, while the convolution stride of the shortcut path Shortcut in the residual downsampling modules (C1 to C5) between the remaining downsampling nodes (X1 to X6) is 2, indicating downsampling. Of course, the convolution stride can also be set to other values, such as 3, 4, 5, etc., so that the width dimension and height dimension of the intermediate tensor data can be rapidly reduced, in order to reduce the computational amount and the complexity of the network model. It should be noted that when each residual downsampling module between the downsampling nodes at each level performs the ReLU activation function operation, the ReLU activation function parameters are set to be the same. And when each of these residual downsampling modules performs the convolution Conv operation and the normalization BN operation, their convolution parameters and normalization parameters can be set to be the same or different to meet the requirements of parameter learning.

[0082] Further, the number of the residual downsampling modules (C0 to C5) can represent (or reflect) the downsampling level. In this embodiment, the number of the residual downsampling modules (C0 to C5) is 6. Except for one residual downsampling module C0 between the input node X1 and the first-level downsampling node X1 that does not perform downsampling, the residual downsampling modules (C1 to C5) between the remaining downsampling nodes all perform downsampling, which reflects that the downsampling level is 5. Since the target (here the billet) in the image is relatively large, it is necessary to increase the receptive field of the network model so that the network model can extract the global structure information of the billet. In view of this, the number of downsamplings needs to be set relatively large. When the downsampling level is set to 5 levels, it can not only meet the positioning accuracy of the billet but also meet the requirements of real-time calculation of the network model. Of course, in other partial embodiments, the number of the residual downsampling modules can also be other values, such as greater than or equal to 4 and less than or equal to 8; correspondingly, the downsampling level is greater than or equal to 3 and less than or equal to 7. It should be noted that the value range of the downsampling level can be m ≤ N ≤ n, where N is the downsampling level, and m and n are the first threshold and the second threshold of the downsampling level. The first threshold is determined according to the receptive field of the preset network model, and the second threshold is determined according to the receptive field of the preset network model and the size of the sample image.

[0083] Refer to Figure 6 and Figure 8 , Figure 8 is Figure 6 the structural schematic diagram of the residual upsampling module shown. Continuing as described above, a residual upsampling module (UC1 to

[0084] UC5) is provided between each level of upsampling nodes. The residual upsampling module (UC1 to UC5) includes a first branch and a second branch; the first branch is used to perform at least one of the operations of convolution Conv operation, normalization BN operation, and activation function ReLU operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path Shortcut, and the second branch is used to obtain a second branch result to perform a superposition operation, an activation function ReLU operation, and an upsampling Upsample operation with the first branch result. As Figure 8 and Figure 9As shown, in this embodiment, the residual upsampling module (UC1-UC5) includes a first branch and a second branch; the first branch is used to sequentially perform a convolution Conv operation, a normalization BN operation, a ReLU activation function operation, a convolution Conv operation, and a normalization BN operation on the intermediate tensor data to obtain a first branch result; the second branch is a shortcut path, and this shortcut path performs a convolution Conv operation and a normalization BN operation on the intermediate tensor data to obtain a second branch result, which is superimposed with the first branch result, followed by a ReLU activation function operation and an upsampling Upsample operation. Among them, for the second branch of the residual upsampling module (UC1-UC5), that is, the shortcut path, the stride is 1 when performing the convolution Conv operation.

[0085] The network model of this embodiment uses a residual module (including a residual downsampling module and a residual upsampling module). Therefore, compared with the traditional convolution module, the residual module can better solve the problems of gradient explosion and network degradation, and further improve the generalization characteristics of the network model during the process of accelerating network learning.

[0086] Continue to refer to Figure 6 , as described above, additional nodes (X12, X13) are also provided between the last-level upsampling node X11 and the output node X14. In this embodiment, the additional node X12 is connected to the last-level upsampling node X11, and the additional node X13 is respectively connected to the additional node X12 and the output node X14. Among them, the additional tensor data of the additional node X11 is based on the intermediate tensor data of the last-level upsampling node and is obtained by performing any one of the convolution Conv operation, the normalization BN operation, and the ReLU activation function operation. The additional tensor data of the additional node X12 is based on the additional tensor data of the additional node X11 and is obtained by performing any one of the convolution Conv operation and the ReLU activation function operation. When the last-level upsampling node X11 and the additional node X12 perform the convolution operation respectively, the convolution operation uses a 3×3 convolution kernel, the stride is 1, and the padding is 1. It should be noted that when the above intermediate nodes (including downsampling nodes, upsampling nodes, and additional nodes) perform the convolution operation, their input and output channel numbers are both determined according to the channel number dimension of the output node X14.

[0087] Continuing as described above, the output node X14 is used to obtain the output tensor data. In this embodiment, the output tensor data is obtained based on the additional tensor data of the additional node X13 and by performing the adaptive average pooling AP operation. Among them, the adaptive average pooling operation AP refers to taking the average of the data of W multiplied by H (specifically, taking the average along the column direction) to obtain the target output of W multiplied by 1. Since the essence of the billet positioning control before entering the furnace is a segmentation task in the horizontal direction, that is, the output node of the network model outputs a one-dimensional signal, it is necessary to perform the adaptive average pooling operation (this operation can be completed by the adaptive average pooling layer) before the output node, that is, project the intermediate output signal of the network model onto the horizontal direction as the final segmentation result of the network model in the horizontal direction. In addition, in this embodiment, the output node also performs the mean squared error loss Loss operation to reduce the error and improve the accuracy of the target output.

[0088] It should be noted that in order to determine the position of the billet in the training sample and only need to determine the longitudinal position of the billet without determining the lateral position of the billet, the output node of the network model outputs the training result, and the training result is a one-dimensional signal. Exemplarily, the one-dimensional signal may include an output vector of length W. In this way, the length of the output vector is the same as the width of the grayscale image input to the network model, and there is a one-to-one correspondence in the longitudinal direction. Further, determine the region position in the output vector that is greater than the preset value. When it is determined that the value corresponding to a certain region position is greater than the preset value, record the left and right end positions (i.e., the predicted label) of the region position as the two end positions of the sample billet. Among them, the above preset value is related to the sample billet and is used to characterize the sample billet. Of course, the one-dimensional signal may also include an output vector whose length is not W. After obtaining the output vector, perform a linear mapping operation. Specifically, map the two end positions representing the sample billet in the output vector to the longitudinal position in the sample original image. That is, divide the length W1 of the sample original image by the output vector W2 to obtain the coefficient w, and then multiply the coefficient w by the left and right end positions of the obtained region position in the output vector, so that the position of the sample billet in the sample original image can be obtained.

[0089] In addition to using the preset network model in step S532 of this embodiment to obtain the training result (which includes the predicted label of the sample billet), it is also possible to implement "training the preset network model based on the training sample to obtain the training result (which includes the predicted label of the sample billet)" by using other artificial neural network models. Exemplarily, other artificial neural network models may be the AlexNet network model, the VGGNet network model, and the ResNet network model.

[0090] Specifically, the network structure of the AlexNet network model includes multiple convolutional stages and fully connected stages. In the convolutional stage, in addition to using convolution, the AlexNet network model also adopts the ReLU activation function, ReLU(x) = max(x, 0), and LRN (Local Response Normalization). Drawing on the idea of lateral inhibition to achieve local inhibition, it makes the values with relatively larger responses even larger, improving the generalization ability of the network model. LRN only normalizes the data in adjacent regions without changing the size and dimension of the data. In addition, the AlexNet network model also applies Overlapping (overlapping pooling), that is, the pooling operation overlaps on some pixels. The size of the pooling kernel is n×n and the stride is k. If k = n, it is normal pooling; if k < n, it is overlapping pooling. In the fully connected stage, the AlexNet network model introduces the function of dropout. Dropout means that during the training process of the network model, for neural network units, they are temporarily discarded from the network (while retaining their weights) according to a certain probability (usually 50%, in which case the randomly generated network structure is the most) and no longer respond to the data transmitted forward and backward. Dropout can effectively prevent the network model from overfitting, making the network have stronger generalization ability. At the same time, due to reducing the complexity of the network model, the operation speed is accelerated.

[0091] Specifically, the network structure of the VGGNet network model performs multiple convolutional operations, max pooling, and activation function operations on the input image, and also performs a fully connected operation. By using the fully connected operation, the output can be flexibly adjusted.

[0092] Specifically, the network structure of the ResNet network model includes direct connection channels. For example, in the ResNet network structure, two types of residual modules are used. One is a series connection of two 3×3 convolutional networks as a residual module, and the other is a series connection of three convolutional networks of 1×1, 3×3, and 1×1 as a residual module. ResNet has different numbers of network layers, and the commonly used ones are 50-layer, 101-layer, and 152-layer. They are all implemented by stacking the above-mentioned residual modules. By directly bypassing the input information to the output to protect the integrity of the information, the entire network only needs to learn the part of the difference between the input and output, thus simplifying the learning objective and difficulty.

[0093] Continue to refer to Figure 5 As shown, in step S534, compare the predicted label and the actual label of the sample billet and obtain the comparison result.

[0094] After the steps of training a preset network model based on training samples to obtain a training result and extracting a predicted label of a sample billet based on the training result, the predicted label and the actual label of the sample billet can be compared. The specific comparison method can be to calculate the error between the actual label and the predicted label. By calculating the error between the two, an error value can be obtained. Further, it is determined whether the error value is within a preset range. When it is determined that the error value exceeds the preset range, the following step S535 is executed.

[0095] Step S535: Adjust the model parameters of the preset network model based on the comparison result until the preset network model converges to obtain a trained network model.

[0096] Based on the comparison result (i.e., the obtained error value above), and by using the gradient descent algorithm to adjust the model parameters of the preset network model until the predicted label and the actual label tend to coincide (i.e., the predicted value tends to the actual value). At this time, the preset network model converges and a trained network model is obtained. It should be noted that the gradient descent algorithm has the characteristics of simple implementation, strong versatility, and good effect. Of course, in other embodiments, methods such as particle swarm, simulated annealing, and surrogate priority can also be used to adjust the model parameters of the preset network model. When the predicted value output by the network model is more and more close to the actual value, that is, the error value is smaller, the training accuracy of the network model can be improved more. In this way, the trained network model can meet the requirement of realizing the accurate positioning of the billet.

[0097] The billet pre-furnace positioning control method of the present application is implemented through the steps of the above method, so as to realize the accurate positioning control of billets with different sizes before entering the furnace, and also avoid affecting the steel charging rhythm due to manual judgment of positioning, and thus can also improve production efficiency.

[0098] To facilitate better implementation of the billet pre-furnace positioning control method provided by the embodiments of the present application, the embodiments of the present application also provide a billet pre-furnace positioning control device based on the above billet pre-furnace positioning control method. The meanings of the nouns are the same as those in the above billet pre-furnace positioning control method, and the specific implementation details can be referred to the description in the method embodiments.

[0099] Please refer to Figure 11 , Figure 11 which is a structural block diagram of a billet pre-furnace positioning control device provided by an embodiment of the present application. The device includes: an image acquisition module 1100, an image processing module 1200, an information acquisition module 1300, and a billet movement module 1400.

[0100] The image acquisition module 1100 is used to acquire the original image of the target billet before entering the furnace.

[0101] For example, the image acquisition module 1100 can be specifically configured to acquire the original image of the target billet before it enters the furnace. The original image can be obtained by using an industrial camera to take pictures.

[0102] The image processing module 1200 is used to perform image processing on the original image to obtain a processed image.

[0103] For example, the image processing module 1200 can be specifically configured to perform image processing on the original image, such as image grayscale processing, but is not limited to image grayscale processing, and can also include image sharpening, image noise reduction, etc.

[0104] The information acquisition module 1300 is used to input the processed image into the trained network model to obtain the position information of the target billet.

[0105] For example, the information acquisition module 1300 can be specifically configured to input the processed image into the trained network model. The trained network model outputs the label of the target billet, and the label includes the labels of both ends of the target billet (i.e., the start-end label and the end-end label). Then, an operation of information extraction can be performed on the label to obtain the position information of the target billet. Specifically, the operation of information extraction can be to first obtain the start-end label and the end-end label of the target billet, and then project the start-end label and the end-end label onto a preset calibration grid, and the position information of the start end of the target billet and the position information of the end end of the target billet can be obtained, that is, the position information of the target billet. Further, according to the position information of the start end and the end end of the target billet, the length information of the target billet can be calculated.

[0106] The billet moving module 1400 is used to adjust the motion state of the roller table for carrying the target billet based on the position information of the target billet and the preset specified position information, and accordingly control the target billet to move to the specified position.

[0107] For example, the billet moving module 1400 can be specifically configured to obtain the position offset between the two positions according to the position information of the target billet obtained by the information acquisition module 1300 and the preset specified position information, and transmit the position offset to the field controller. The field controller can control the motion state of the roller table for carrying the target billet, and through the control of the field controller, the target billet is moved to the specified position.

[0108] By using the above-mentioned modules in cooperation, the billet pre-furnace positioning control device 1000 of the present application can achieve precise positioning control of billets with different sizes before they enter the furnace, and also avoid affecting the steel charging rhythm due to manual judgment of positioning, thus improving production efficiency.

[0109] See Figure 12 , Figure 12The structural block diagram of the billet positioning control device 1000 provided by another embodiment of the present application. In another embodiment of the present application, in addition to the image acquisition module 1100, the image processing module 1200, the information acquisition module 1300, and the billet moving module 1400 in the above embodiment, the device 1000 may further include: a sample image acquisition module 1500, a sample image calibration module 1600, and a sample image training module 1700. Among them, the sample image acquisition module 1500, the sample image calibration module 1600, and the sample image training module 1700 are connected in sequence, and the sample image training module 1700 is connected to the information acquisition module 1300.

[0110] The sample image acquisition module 1500 is configured to acquire the original sample image of the sample billet before entering the furnace.

[0111] For example, the sample image acquisition module 1500 may specifically be configured to collect a large number of sample images generated throughout a day or longer when different steel grades are generated, so as to meet the quantity requirements of the samples, and thus be able to better train the network model. It should be noted that the sample image acquisition module 1500 can obtain the actual sample images in the on-site production process through preset cameras, light sources, and other network devices. Among them, high-quality lenses can be selected for the cameras, and high-performance configured hardware can be used for the network devices, so that the sample image acquisition module 1500 can obtain high-quality sample images to meet the quality requirements of the samples and further improve the training accuracy of the network model.

[0112] The sample image calibration module 1600 is configured to perform calibration processing on the original sample image to obtain the calibrated sample image and the actual label of the sample billet. In a further embodiment of the present application, the sample image calibration module 1600 may also be configured to perform image preprocessing on the calibrated sample image. The preprocessing may include image grayscale conversion. Image grayscale conversion is to set the pixel points on the image to a sampled color. For example, the calibrated sample image (which is a color image) can be converted into a grayscale image through the following formula. The formula is Gray = R * 0.299 + G * 0.587 + B * 0.114, where Gray represents grayscale, R represents red pixels, G represents green pixels, and B represents blue pixels. In this way, through the image preprocessing of the calibrated sample image, the edge processing of the image is effectively realized, and the expressiveness of the image is improved. In addition, the preprocessing may also include image noise reduction and image sharpening. Exemplarily, the image sharpening is Laplacian operator image sharpening. After the image grayscale conversion, the Laplacian operator image sharpening is performed on the grayscale image to highlight the obvious contour curves in the image and hide the relatively unobvious image boundaries. Specifically, the Laplacian sharpened image is related to the degree of mutation of the surrounding pixels of a certain pixel in the image to this pixel, that is, the basis of the Laplacian sharpened image is the degree of change of the image pixels.

[0113] Continue to refer to Figure 11 Figure 11 , the sample image training module 1700 is used to train a preset network model based on the calibrated sample image and the actual label to obtain a trained network model. Refer to in combination Figure 12 , Figure 12 is Figure 11 The structural block diagram of the sample image training module shown. In a further embodiment of the present application, the sample image training module 1700 may include: a model parameter initialization unit 1710, a training sample generation unit 1720, a label information acquisition unit 1730, a label information comparison unit 1740, and a model parameter adjustment unit 1750.

[0114] The model parameter initialization unit 1710 is used to initialize the model parameters of the preset network model.

[0115] The training sample generation unit 1720 is used to generate training samples according to the calibrated sample image.

[0116] The label information acquisition unit 1730 is used to train the preset network model based on the training samples to obtain a training result, and extract the predicted label of the sample billet based on the training result. Among them, the preset network model is a deep learning network model. The deep learning network model includes an input node, an intermediate node, and an output node, where the input node is used to receive training samples to obtain input tensor data; the output node is used to obtain output tensor data and output a training result accordingly, where the training result includes the predicted label of the sample billet.

[0117] The label information comparison unit 1740 compares the predicted label of the sample billet with the actual label and obtains a comparison result. Specifically, the label information comparison unit 1740 can be used to calculate the error between the actual label and the predicted label and obtain an error value. Further, it is determined whether the error value is within a preset range. When it is determined that the error value exceeds the preset range, the model parameter adjustment unit is called.

[0118] The model parameter adjustment unit 1750 adjusts the model parameters of the preset network model based on the comparison result until the preset network model converges to obtain a trained network model. Specifically, based on the comparison result, this unit can adjust the model parameters of the preset network model through the gradient descent algorithm until the predicted label and the actual label tend to coincide (that is, the predicted value tends to the actual value). At this time, the preset network model converges and a trained network model is obtained. When the predicted value output by the network model is more inclined to the actual value, that is, the error value is smaller, the training accuracy of the network model can be improved more. In this way, the trained network model can meet the requirements for realizing the accurate positioning of the billet.

[0119] Before the billet enters the furnace, the positioning control device 1000 described in this application can achieve precise positioning control of billets with different sizes before entering the furnace through the cooperation of the above modules or units (including sub-units), and also avoid affecting the steel charging rhythm due to manual judgment of positioning, thereby improving production efficiency.

[0120] In addition, in an embodiment of this application, an electronic device 5000 is also provided, as Figure 13 shown. The electronic device 5000 may include at least one processor 5100 and at least one memory 5200. Those skilled in the art can understand that Figure 13 the electronic device 5000 shown in does not constitute a limitation on the electronic device, and may include more or fewer components than shown, or combine certain components, or have different component arrangements. Among them:

[0121] The processor 5100 is the control center of the electronic device 5000. By running or executing software programs and / or modules stored in the memory 5200, and calling data stored in the memory 5200, it executes various functions of the electronic device 5000 and processes data, thereby monitoring the electronic device 5000 as a whole. Optionally, the processor 5100 may include one or more processing cores; preferably, the processor 5100 may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, user interface, application programs, etc., and the modem processor mainly processes wireless communication. It can be understood that the above modem processor may not be integrated into the processor 5100.

[0122] The memory 5200 can be used to store software programs and modules. The processor 5100 executes various functional applications and data processing by running the software programs and modules stored in the memory 5200 to achieve various functions, specifically as follows:

[0123] Obtain the original image of the target billet before entering the furnace;

[0124] Perform image processing on the original image to obtain a processed image;

[0125] Input the processed image into the trained network model to obtain the position information of the target billet; and

[0126] Based on the position information of the target billet and the preset specified position information, adjust the movement state of the roller table for carrying the target billet, and accordingly control the target billet to move to the specified position.

[0127] Those of ordinary skill in the art can understand that all or part of the steps in the methods described in the above embodiments can be completed by instructions or by controlling relevant hardware through instructions. These instructions can be stored in a computer-readable storage medium and loaded and executed by the processor 5100.

[0128] Therefore, an embodiment of the present application provides a computer-readable storage medium, in which multiple computer programs are stored. These computer programs can be loaded by a processor to execute the steps in a billet positioning control method before entering the furnace provided in any embodiment of the present application. For example, the computer program can execute the following steps:

[0129] Obtain the original image of the target billet before entering the furnace;

[0130] Perform image processing on the original image to obtain a processed image;

[0131] Input the processed image into the trained network model to obtain the position information of the target billet; and

[0132] Based on the position information of the target billet and the preset specified position information, adjust the movement state of the roller table for carrying the target billet, and accordingly control the target billet to move to the specified position.

[0133] For the specific implementation of each of the above operations, reference can be made to the previous embodiments, which will not be elaborated here. Among them, the storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, etc.

[0134] Since the instructions stored in the computer-readable storage medium can execute the steps in a billet positioning control method before entering the furnace provided in any embodiment of the present application, the beneficial effects achievable by a billet positioning control method before entering the furnace provided in any embodiment of the present application can be realized. For details, refer to the previous embodiments, which will not be elaborated here.

[0135] The above has introduced in detail a billet positioning control method, device, electronic device, and computer-readable storage medium provided in an embodiment of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the technical solution and its core idea of the present application; those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A positioning control method for a steel billet before entering a furnace, characterized in that, Including: Obtaining the original image of the target billet before it enters the furnace; Performing image processing on the original image to obtain a processed image; Inputting the processed image into the trained network model to obtain the position information of the target billet; And Based on the position information of the target billet and the preset specified position information, adjusting the motion state of the roller table for carrying the target billet, and accordingly controlling the target billet to move to the specified position; Before the step of inputting the processed image into the trained network model, it includes: Collecting the sample original image of the sample billet before it enters the furnace; Performing calibration processing on the sample original image to obtain a calibrated sample image and the actual label of the sample billet; Based on the calibrated sample image and the actual label, training the preset network model to obtain a trained network model; The preset network model includes an input node, an intermediate node, and an output node. The input node is used to receive training samples to obtain input tensor data; the output node is used to obtain output tensor data and output training results accordingly; the intermediate node includes multiple levels of downsampling nodes and multiple levels of upsampling nodes; each level of downsampling node is connected to the next level of downsampling node through a corresponding residual downsampling module. Each level of downsampling node is used to obtain the intermediate tensor data of this level of downsampling node according to the input tensor data and the intermediate tensor data of the previous level of downsampling node; each level of upsampling node is connected to the next level of upsampling node through a corresponding residual upsampling module. Each level of upsampling node is used to obtain the intermediate tensor data of this level of upsampling node according to the intermediate tensor data of the last level of downsampling node and the intermediate tensor data of the previous level of upsampling node; each level of downsampling node is cross-connected to the corresponding upsampling node.

2. The method according to claim 1, wherein The step of performing calibration processing on the sample original image to obtain a calibrated sample image and the actual label of the sample billet includes: Performing image preprocessing on the calibrated sample image.

3. The method according to claim 1, wherein The step of training the preset network model based on the calibrated sample image and the actual label to obtain a trained network model includes: Initializing the model parameters of the preset network model; Generating training samples according to the calibrated sample image; Training the preset network model based on the training samples to obtain training results, and extracting the predicted label of the sample billet based on the training results; Comparing the predicted label and the actual label of the sample billet and obtaining a comparison result; Adjusting the model parameters of the preset network model based on the comparison result until the preset network model converges to obtain a trained network model.

4. The method according to claim 1, wherein The number of the residual downsampling modules is used to represent the number of downsampling levels. The value range of the number of downsampling levels is m ≤ N ≤ n, where N is the number of downsampling levels, and m and n are the first threshold and the second threshold of the number of downsampling levels. The first threshold is determined according to the receptive field of the preset network model, and the second threshold is determined according to the receptive field of the preset network model and the size of the sample image.

5. The method according to claim 1, wherein An additional node is further provided between the last-level upsampling node and the output node, and the additional tensor data of the additional node is obtained based on the intermediate tensor data of the last-level upsampling node and by performing at least one of a convolution operation, a normalization operation, and an activation function operation.

6. The method according to claim 5, characterized in that, The output tensor data of the output node is obtained based on the additional tensor data and by performing an adaptive average pooling operation.

7. The method according to claim 1, wherein The residual downsampling module includes a first branch and a second branch; the first branch is configured to perform at least one of a convolution operation, a normalization operation, and an activation function operation on the intermediate tensor data to obtain a first branch result; The second branch is a shortcut path, and the second branch is configured to obtain a second branch result for performing a superposition operation and an activation function operation with the first branch result.

8. The method according to claim 1, wherein The residual upsampling module includes a first branch and a second branch; the first branch is configured to perform at least one of a convolution operation, a normalization operation, and an activation function operation on the intermediate tensor data to obtain a first branch result; The second branch is a shortcut path, and the second branch is configured to obtain a second branch result for performing a superposition operation, an activation function operation, and an upsampling operation with the first branch result.

9. The method according to claim 7 or 8, characterized in that The shortcut path is configured to perform a convolution operation and a normalization operation on the intermediate tensor data.

10. A positioning control device for a steel billet before entering a furnace, characterized in that, Controlled by using the billet pre-furnace positioning control method according to any one of claims 1-9, the device includes: An image acquisition module, configured to acquire an original image of a target billet before entering the furnace; An image processing module, configured to perform image processing on the original image to obtain a processed image; An information acquisition module, configured to input the processed image into a trained network model to obtain the position information of the target billet; and A billet movement module, configured to adjust the movement state of a roller table for carrying the target billet based on the position information of the target billet and preset specified position information, thereby controlling the target billet to move to the specified position.

11. An electronic device, characterized in that, Including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the steps of the billet pre-furnace positioning control method according to any one of claims 1 to 9 are implemented.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, the steps of the billet pre-furnace positioning control method according to any one of claims 1 to 9 are implemented.

Images