A molten iron information corresponding method and system based on visual recognition

Abstract

The present application relates to a kind of based on visual identification molten iron information corresponding method and system, belong to metallurgical method and system technical field.The technical scheme of the present application is: molten iron ladle position visual identification unit will be identified molten iron ladle number and sampling time transmission to molten iron information integration transmission unit, laboratory analysis system will temperature, ladle number and composition information corresponding upload to molten iron information integration transmission unit, converter iron position visual identification unit will be molten iron ladle number, molten iron weight and iron time information transmission to molten iron information integration transmission unit, molten iron information integration transmission unit will comprehensive information transmission to converter secondary system for converter secondary model calculation.The beneficial effects of the present application are: realize the data processing mode of whole process without manual intervention, improve the accuracy of main data of converter smelting, reduce the manual determination process, improve the accuracy of converter automation and model production.

Description

TECHNICAL FIELD

[0001] The present application relates to a molten iron information corresponding method and system based on visual recognition, belonging to the technical field of metallurgical methods and systems. BACKGROUND

[0002] Molten iron is the main raw material for converter smelting, and the molten iron charging information mainly includes molten iron composition information, molten iron temperature information and molten iron weight information. The above information is the main process parameter for converter smelting model calculation and is the basic data for converter static model calculation, directly affecting the converter smelting heat balance and material balance calculation, affecting the converter slag making calculation, and having important significance for improving the stability of the overall converter smelting process and ensuring the quality of the final product.

[0003] In the traditional production process, there are differences in time and space between the molten iron pouring station and the converter iron mixing, including the differences in time and space of the appearance of the molten iron information, such as the molten iron temperature information obtained after pouring, the molten iron composition information obtained by laboratory analysis after sampling from the molten iron ladle, and the molten iron weight information obtained by the crane weighing after the converter iron mixing. The molten iron process goes to different places, such as being directly poured into the converter for smelting or being poured into the molten iron pretreatment process for desulfurization. There is also complexity in the cross operation of the crane, such as the problem caused by the molten iron desulfurization or production rhythm, the cross hoisting of the molten iron ladle by different cranes, and the uncertainty of production scheduling, which cannot be matched in advance. Therefore, in the process of integrating and corresponding the molten iron information, it is impossible to realize the corresponding and accurate transmission of the molten iron information through simple logic or mechanical language.

[0004] In most converter smelting systems, the identification and tracking of molten iron information mainly rely on manual confirmation, such as the Daniele second-level system, which combines the chemical composition of the molten iron ladle number information with the steelmaking laboratory to judge and integrate the temperature and composition information of the molten iron used for smelting by the post operator. However, due to the time and space separation of the molten iron pouring position, the laboratory and the converter operating position, it is easy to cause deviation in information communication. The molten iron weight information is calculated by the swing furnace worker by observing the electronic display weight before and after the crane weighing, and the difference is manually calculated. This method also has the problem of calculation error.

[0005] Currently, there is no public feasible method for corresponding and transmitting the molten iron information.

[0006] The patent with application number 202010355970.9 discloses a "system and method for identifying iron and steel ladle number by using infrared visual recognition", which installs a target number on the support connected to the outer wall surface of the iron and steel ladle, mainly a specific symbol formed by punching, and obtains an infrared image by real-time collection of the target number through an infrared thermal imaging camera and a lens to identify the iron and steel ladle number. This method identifies the iron and steel ladle number, but does not realize the correspondence of molten iron information, and the method needs to punch the target number, increasing the workload on site.

[0007] The patent with application number 201910486497.5 discloses a "torpedo ladle visual recognition and positioning method and system for metallurgical operation", which mainly includes coarse positioning to obtain a region of interest by sliding window template matching of the torpedo ladle opening, and the visual recognition and positioning algorithm performs image preprocessing and target search in the region of interest to complete fine positioning and feature fitting. The target feature extraction algorithm uses a parameter adaptive ellipse feature detection algorithm. However, this method is mainly for identification and positioning of molten iron torpedo ladle, and cannot effectively transmit the molten iron information in the torpedo ladle.

[0008] The patent with application number 202210252269.3 discloses a "method for real-time tracking of iron ladle and steel ladle", which installs RFID identification tags on all iron ladles, steel ladles and molten iron ladle trucks, sets identification codes, and determines the entry of iron ladle into the ladle position of the steelmaking plant through track laser positioning, molten iron ladle RFID and iron ladle IFID identification. This invention only identifies and positions the iron ladle and steel ladle, and cannot effectively transmit the molten iron information.

[0009] Therefore, there is a need to invent a molten iron information correspondence method to improve the accuracy of the molten iron information for the converter, improve the accuracy of the automatic and model production of the converter, and reduce the labor intensity of the operator. SUMMARY

[0010] The purpose of the present application is to provide a molten iron information correspondence method and system based on visual recognition, which establishes a correspondence method between molten iron information and raw material information required for converter smelting by visual recognition method, realizes a full-process data processing mode without manual intervention, improves the accuracy of main data for converter smelting, reduces the manual determination process, improves the accuracy of automatic and model production of the converter, effectively avoids the composition accidents, smoking environmental protection and splashing safety accidents caused by inaccurate raw material information, lays a foundation for the green and intelligent transformation and upgrading of steel enterprises, and effectively solves the above problems existing in the background technology.

[0011] The technical scheme of the present application is: a molten iron information correspondence method based on visual recognition, comprising the following steps:

[0012] (1) Hot metal is transported to the steelmaking plant by torpedo ladle, and hot metal turning operation is performed. After the hot metal turning operation is completed, the ladle is opened by the ground car, temperature sampling is performed, the visual recognition unit collects the ladle number and hoisting time, and the recognized ladle number and hoisting time are transmitted to the hot metal information integration transmission unit. The hot metal sample is sent to the laboratory for detection, and the temperature and corresponding ladle number information are transmitted to the laboratory analysis system. After the laboratory composition analysis is completed, the temperature, ladle number and composition information are uploaded to the hot metal information integration transmission unit.

[0013] The hot metal ladle hoisting process is recognized by the industrial camera of the hot metal ladle position visual recognition unit. The running state of the ladle is automatically recognized by the lifting and ladle return recognition logic. The number on the ladle identification is automatically recognized by the picture processing program and the deep learning artificial intelligence model.

[0014] (2) The ladle is hoisted to the converter front, and the ladle number and crane scale weight during the ladle-to-converter stage are recognized by the converter iron position visual recognition unit. The ladle number, hot metal weight and ironing time information are transmitted to the hot metal information integration transmission unit.

[0015] The ironing state is determined by the industrial visual recognition camera and the converter ironing logic of the ironing position visual recognition unit. The number on the ladle identification is automatically recognized by the picture processing program and the deep learning artificial intelligence model.

[0016] (3) The hot metal information integration transmission unit analyzes and integrates the information provided by the previous units of the system by computer. The hot metal composition and temperature logic and the hot metal weight logic are used to finally transmit the comprehensive information of the hot metal composition, temperature and weight used by the converter to the converter secondary system for converter secondary model calculation.

[0017] In step (1), the lifting and ladle return recognition logic sets the recognized sensitive area as a vertical rectangle, which is divided into two parts from the center, the upper part is 1 area and the lower part is 2 area. When the ladle number recognition digits appear in 2 area and 1 area in turn, it is the lifting process of the ladle after turning, at this time, the ladle number and hoisting time are recorded. 吊 (n) is the hoisting time, where n is the ladle number. When the ladle number recognition digits appear in 1 area and 2 area in turn, it is the empty ladle return process of the ladle, at this time, no data recording is required.

[0018] In step (1), the picture processing program considers improving the calculation efficiency and the ladle number recognition speed when setting the sensitive area. Once the ladle is recognized in the photo and the ladle has reached the specified position, the ladle photo will be:

[0019] ① Grayscale processing: First, the photo is inverted to grayscale. The value of all pixels in the image is 255 - the current pixel value.

[0020] ② Binarization processing: Gaussian filtering is performed on all pixels of the image in 3*3 pixel units;

[0021] ③ The image pixels are eroded in 3*3 pixel units, and the value of the processed pixel is reduced to the minimum value of its neighboring pixels;

[0022] ④ Perform adaptive filtering on the image. The pixel value of the processed pixel = the average value of adjacent pixels in a 3*3 pixel unit - a constant, where the constant is defined as 10.

[0023] ⑤ The edge-finding process is used to find the target number;

[0024] ⑥ Perform contour processing to select the identified digits. Extract the digits from the sensitive area.

[0025] In step (1), the key steps for setting up the deep learning artificial intelligence model are as follows:

[0026] (a) A large number of photos of ladle numbers were taken using an industrial camera, and these photos were manually labeled to form a training set of labels;

[0027] (b) The dataset is divided into a training set and a test set, with the training set accounting for 70% of the total number of photos and the test set accounting for 30% of the total number of photos;

[0028] (c) Organize the training set into a (42, 42, 1) format according to the length, width and grayscale of the images. The matrix format of the images is a two-dimensional matrix with 42 rows, 42 columns and 1 layer.

[0029] (d) Use the formula X_train / = 255 # to normalize the pixel values ​​of the training set data to between 0 and 1. This is a matrix division, dividing 255 by each element in the matrix.

[0030] (e) Convert the labels into one-hot encoded form to facilitate the training of the model for multi-class classification tasks; it represents each category as a binary vector, in which only one element is 1 and the rest are 0; specifically, for a classification problem with n categories, each category is represented by a vector of length n, the i-th element of the vector is 1, indicating that the vector represents the i-th category, while the other elements are 0.

[0031] (f) Establish a sequential model to facilitate the design of deep learning neural network architecture;

[0032] (g) Add a convolutional layer with 32 kernels, a kernel size of 3x3, and the activation function ReLU. The input data shape is (28, 28, 1).

[0033] (h) Add another convolutional layer with 64 kernels, a kernel size of 3x3, and the activation function ReLU;

[0034] (i) Add a 2x2 max pooling layer;

[0035] (j) Add a dropout layer with a dropout rate of 25%;

[0036] (k) Flatten the data into 1-dimensional data;

[0037] (l) Add a fully connected layer with 128 nodes and ReLU activation function;

[0038] (m) Add a dropout layer with a dropout rate of 50%;

[0039] (n) Add an output layer with 10 nodes and softmax activation function;

[0040] (o) Compilation model, loss function is cross-entropy, optimizer is Adam, and evaluation metric is accuracy.

[0041] In step (2), the converter molten iron charging logic refers to identifying the ladle number information during the molten iron charging process. Its sensitive area is a long, flat rectangle, with area 1 being closer to the converter and area 2 being further away. When the ladle number appears successively in areas 2 and 1, the converter begins the molten iron charging process. At this time, the ladle number information, time information, and overhead crane weighing information are recorded using a deep learning artificial intelligence model. T兑 (n) represents the weight of the ladle before the start of molten iron charging, Tcharging represents the start time of molten iron charging, and n represents the ladle number. When the ladle number appears in zone 1 and zone 2 respectively, the converter molten iron charging process ends. At this time, the ladle number information, time information, and overhead crane weighing information are recorded. W T结 (n) represents the weight of the molten iron ladle at the end of the molten iron exchange, T represents the end time of the molten iron exchange, and n represents the ladle number;

[0042] The image processing program and the deep learning artificial intelligence model are the same as those used for the can-turning position.

[0043] In step (3), the logic for the composition and temperature of molten iron is as follows: the time relationship between the ladle number and the corresponding ladle number in the laboratory is T. 吊 (n) ≤ T 验 (n) ≤ T 兑 At time (n), the chemical composition and temperature corresponding to iron ladle number n are T. 验 (n) Information corresponding to time T, where T吊 (n) represents the time T is spent in the flipping position of the iron bag with bag number n. 兑 (n) represents the time T is spent at the iron exchange point for iron package number n. 验 (n) represents the time it takes for the iron bag with package number n to be analyzed in the laboratory;

[0044] The weight logic for molten iron is: W 铁 (n) = W T兑 (n)-W T结 (n), where W 铁 (n) represents the weight of the molten iron in ladle number n, W T兑 (n) represents the weight of the molten iron before the start of the iron-making process, W. T结 (n) represents the weight of the molten iron after the iron-making process is completed.

[0045] A visual recognition-based molten iron information correspondence system includes a laboratory analysis system, a molten iron pouring position visual recognition unit, a converter iron-adding position visual recognition unit, and a molten iron information integration and transmission unit. The laboratory analysis system, the molten iron pouring position visual recognition unit, and the converter iron-adding position visual recognition unit are respectively connected to the input end of the molten iron information integration and transmission unit, and the output end of the molten iron information integration and transmission unit is connected to the converter secondary system.

[0046] The molten iron pouring position visual recognition unit includes an industrial camera, a lifting and repacking recognition logic unit, an image processing program, and a deep learning artificial intelligence model. The lifting and repacking recognition logic unit, the image processing program, and the deep learning artificial intelligence model are respectively connected to the industrial camera.

[0047] The converter iron-feeding position visual recognition unit includes an industrial camera, a converter iron-feeding logic unit, an image processing program, and a deep learning artificial intelligence model. The converter iron-feeding logic unit, the image processing program, and the deep learning artificial intelligence model are respectively connected to the industrial camera.

[0048] The beneficial effects of this invention are: by establishing a method for corresponding molten iron information with raw material information required for converter smelting through visual recognition, a data processing mode without human intervention is realized, the accuracy of key data in converter smelting is improved, the manual judgment process is reduced, the accuracy of converter automation and model-based production is improved, and composition accidents, smoke and environmental pollution, and splashing safety accidents caused by inaccurate raw material information in converter smelting are effectively avoided, laying the foundation for the green and intelligent transformation and upgrading of steel enterprises. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the system functions of the present invention.

[0050] Figure 2 This is a diagram showing the layout of the industrial camera in this invention;

[0051] Figure 3 This is a diagram of the sensitive areas identified by the visual recognition unit for molten iron pouring positions in this invention.

[0052] Figure 4 This is a diagram of the sensitive area identified by the visual recognition unit for molten iron mixing position in this invention.

[0053] Figure 5 This is a diagram for identifying the ladle number of molten iron in the ladle according to the present invention;

[0054] Figure 6 This is a grayscale processing result image of the present invention;

[0055] Figure 7 This is a diagram showing the binarization result of the present invention;

[0056] Figure 8 This is a diagram showing the results of the molten iron ladle number identification according to the present invention;

[0057] Figure 9 This is a diagram for identifying the ladle number when adding molten iron to the ladle, as per the present invention.

[0058] Figure 10 This is a diagram for identifying the weight of molten iron in this invention. Implementation

[0059] To make the purpose, technical solutions, and advantages of the invention's embodiments clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described are only a small part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0060] A method for mapping molten iron information based on visual recognition includes the following steps:

[0061] (1) Molten iron is transported from the torpedo ladle to the ladle-turning station of the steelmaking plant for molten iron turning operations. After the turning operations are completed, the molten iron ladle is driven out by the ground car for temperature measurement and sampling. The ladle number and hoisting time are collected by the visual recognition unit and the identified molten iron ladle number and hoisting time are transmitted to the molten iron information integration and transmission unit. The molten iron sample is sent to the testing station for testing. The temperature and corresponding molten iron ladle number information are transmitted to the laboratory analysis system. After the laboratory composition analysis is completed, the temperature, ladle number and composition information are uploaded to the molten iron information integration and transmission unit accordingly.

[0062] The molten iron ladle pouring visual recognition unit uses an industrial camera to identify the process of hoisting the molten iron ladle out, automatically identifies the running status of the molten iron ladle through hoisting and return recognition logic, and automatically recognizes the numbers marked on the molten iron ladle through image processing program and deep learning artificial intelligence model.

[0063] (2) The molten iron ladle is hoisted to the front of the converter furnace. The molten iron ladle number and the weight weighed by the crane scale are identified by the visual recognition unit of the converter iron-pouring position. The molten iron ladle number, molten iron weight and iron-pouring time information are transmitted to the molten iron information integration and transmission unit.

[0064] The iron-feeding position visual recognition unit uses an industrial visual recognition camera to determine the iron-feeding status using the converter iron-feeding logic, and automatically recognizes the numbers marked on the molten iron ladle through an image processing program and a deep learning artificial intelligence model.

[0065] (3) The molten iron information integration and transmission unit analyzes and integrates the information provided by the previous units of the system through the computer, and uses the molten iron composition and temperature logic and molten iron weight logic to finally transmit the comprehensive information of molten iron composition, temperature and weight used in the converter to the converter secondary system for converter secondary model calculation.

[0066] In step (1), the lifting and return identification logic is as follows: the sensitive area for identification is set as a vertically elongated rectangle, which is divided into upper and lower parts from the center. The upper part is area 1, and the lower part is area 2. When the ladle number identification number appears in area 2 and area 1 respectively, it indicates the lifting process after the molten iron ladle is turned over. At this time, the molten iron ladle number and the lifting time are recorded. 吊 (n) represents the hoisting time, where n is the ladle number. When the ladle number identification digits appear in zone 1 and zone 2 respectively, it is the process of returning the empty ladle to its original position. At this time, no data recording is required.

[0067] In step (1), the image processing program, while setting the sensitive area, also considers improving computational efficiency and increasing the speed of ladle number recognition. Once it is recognized that there is a ladle in the photo and the ladle has reached the designated position, the ladle photo will be processed as follows:

[0068] ① Grayscale processing: First, the photo is inverted to grayscale. The value of all pixels in the image is 255 - the current pixel value.

[0069] ② Binarization processing: Gaussian filtering is performed on all pixels of the image in 3*3 pixel units;

[0070] ③ The image pixels are eroded in 3*3 pixel units, and the value of the processed pixel is reduced to the minimum value of its neighboring pixels;

[0071] ④ Perform adaptive filtering on the image. The pixel value of the processed pixel = the average value of adjacent pixels in a 3*3 pixel unit - a constant, where the constant is defined as 10.

[0072] ⑤ The edge-finding process is used to find the target number;

[0073] ⑥ Perform contour processing to select the identified digits. Extract the digits from the sensitive area.

[0074] In step (1), the key steps for setting up the deep learning artificial intelligence model are as follows:

[0075] (a) A large number of photos of ladle numbers were taken using an industrial camera, and these photos were manually labeled to form a training set of labels;

[0076] (b) The dataset is divided into a training set and a test set, with the training set accounting for 70% of the total number of photos and the test set accounting for 30% of the total number of photos;

[0077] (c) Organize the training set into a (42, 42, 1) format according to the length, width and grayscale of the images. The matrix format of the images is a two-dimensional matrix with 42 rows, 42 columns and 1 layer.

[0078] (d) Use the formula X_train / = 255 # to normalize the pixel values ​​of the training set data to between 0 and 1. This is a matrix division, dividing 255 by each element in the matrix.

[0079] (e) Convert the labels into one-hot encoded form to facilitate the training of the model for multi-class classification tasks; it represents each category as a binary vector, in which only one element is 1 and the rest are 0; specifically, for a classification problem with n categories, each category is represented by a vector of length n, the i-th element of the vector is 1, indicating that the vector represents the i-th category, while the other elements are 0.

[0080] (f) Establish a sequential model to facilitate the design of deep learning neural network architecture;

[0081] (g) Add a convolutional layer with 32 kernels, a kernel size of 3x3, and the activation function ReLU. The input data shape is (28, 28, 1).

[0082] (h) Add another convolutional layer with 64 kernels, a kernel size of 3x3, and the activation function ReLU;

[0083] (i) Add a 2x2 max pooling layer;

[0084] (j) Add a dropout layer with a dropout rate of 25%;

[0085] (k) Flatten the data into 1-dimensional data;

[0086] (l) Add a fully connected layer with 128 nodes and ReLU activation function;

[0087] (m) Add a dropout layer with a dropout rate of 50%;

[0088] (n) Add an output layer with 10 nodes and softmax activation function;

[0089] (o) Compilation model, loss function is cross-entropy, optimizer is Adam, and evaluation metric is accuracy.

[0090] In step (2), the converter molten iron charging logic refers to identifying the ladle number information during the molten iron charging process. Its sensitive area is a long, flat rectangle, with area 1 being closer to the converter and area 2 being further away. When the ladle number appears successively in areas 2 and 1, the converter begins the molten iron charging process. At this time, the ladle number information, time information, and overhead crane weighing information are recorded using a deep learning artificial intelligence model. T兑 (n) represents the weight of the ladle before the start of molten iron charging, Tcharging represents the start time of molten iron charging, and n represents the ladle number. When the ladle number appears in zone 1 and zone 2 respectively, the converter molten iron charging process ends. At this time, the ladle number information, time information, and overhead crane weighing information are recorded. W T结 (n) represents the weight of the molten iron ladle at the end of the molten iron exchange, T represents the end time of the molten iron exchange, and n represents the ladle number;

[0091] The image processing program and the deep learning artificial intelligence model are the same as those used for the can-turning position.

[0092] In step (3), the logic for the composition and temperature of molten iron is as follows: the time relationship between the ladle number and the corresponding ladle number in the laboratory is T. 吊 (n) ≤ T 验 (n) ≤ T 兑 At time (n), the chemical composition and temperature corresponding to iron ladle number n are T. 验 (n) Information corresponding to time T, where T 吊 (n) represents the time T is spent in the flipping position of the iron bag with bag number n. 兑 (n) represents the time T is spent at the iron exchange point for iron package number n. 验 (n) represents the time it takes for the iron bag with package number n to be analyzed in the laboratory;

[0093] The weight logic for molten iron is: W 铁 (n) = W T兑 (n)-W T结 (n), where W 铁 (n) represents the weight of the molten iron in ladle number n, W T兑 (n) represents the weight of the molten iron before the start of the iron-making process, W. T结 (n) represents the weight of the molten iron after the iron-making process is completed.

[0094] A visual recognition-based molten iron information correspondence system includes a laboratory analysis system, a molten iron pouring position visual recognition unit, a converter iron-adding position visual recognition unit, and a molten iron information integration and transmission unit. The laboratory analysis system, the molten iron pouring position visual recognition unit, and the converter iron-adding position visual recognition unit are respectively connected to the input end of the molten iron information integration and transmission unit, and the output end of the molten iron information integration and transmission unit is connected to the converter secondary system.

[0095] The molten iron pouring position visual recognition unit includes an industrial camera, a lifting and repacking recognition logic unit, an image processing program, and a deep learning artificial intelligence model. The lifting and repacking recognition logic unit, the image processing program, and the deep learning artificial intelligence model are respectively connected to the industrial camera.

[0096] The converter iron-feeding position visual recognition unit includes an industrial camera, a converter iron-feeding logic unit, an image processing program, and a deep learning artificial intelligence model. The converter iron-feeding logic unit, the image processing program, and the deep learning artificial intelligence model are respectively connected to the industrial camera.

[0097] In practical applications, the specific implementation steps are as follows:

[0098] 1. Molten iron is transported from a torpedo ladle to the ladle-turning station in the steelmaking plant for molten iron turning. After the turning operation is completed, the molten iron ladle is driven out by a ground car for temperature measurement and sampling. The ladle number and hoisting time are collected by a visual recognition unit, and the identified ladle number and sampling time are transmitted to the molten iron information integration and transmission unit. The molten iron sample is sent to the testing station for analysis. The temperature and corresponding ladle number information are transmitted to the laboratory analysis system. After the laboratory completes the composition analysis, the temperature, ladle number, and composition information are uploaded to the molten iron information integration and transmission unit.

[0099] The molten iron ladle pouring visual recognition unit mainly uses an industrial camera for visual recognition. The camera recognizes the process of hoisting the molten iron ladle out, and automatically identifies the running status of the molten iron ladle through hoisting and return recognition logic. It also automatically recognizes the numbers marked on the molten iron ladle through image processing program and deep learning artificial intelligence model.

[0100] The lifting and return identification logic is as follows: the sensitive area for identification is set as a vertically elongated rectangle, which is divided into upper and lower parts from the center. The upper part is area 1, and the lower part is area 2. When the ladle number appears successively in area 2 and area 1, it indicates the lifting process after the molten iron ladle is turned over. At this time, the ladle number and the transportation time are recorded. 吊 (n) represents the hoisting time, where n is the ladle number. When the ladle number identification digits appear successively in zone 1 and zone 2, it is the process of returning the empty ladle to its original position, and no data recording is required at this time.

[0101] The image processing program primarily considers improving computational efficiency and increasing the speed of ladle number recognition while setting sensitive areas. Once a ladle is detected in the photo and has reached the designated location, the ladle photo will be processed as follows:

[0102] ① Grayscale processing: First, reverse the grayscale of the photo.

[0103] The total pixel value in the image = 255 - current pixel value

[0104] ② Binarization processing: Gaussian filtering is performed on all pixels of the image in 3*3 pixel units.

[0105] ③ The image pixels are eroded in 3*3 pixel units, and the value of the processed pixel is reduced to the minimum value of its neighboring pixels.

[0106] ④ Perform adaptive filtering on the image. The pixel value of the processed pixel is calculated as the average value of adjacent pixels in a 3x3 pixel unit minus a constant. In this invention, the constant is defined as 10.

[0107] ⑤ Find the target number by tracing the edge.

[0108] ⑥ Perform contour processing to select the identified digits. Extract the digits from the sensitive area.

[0109] The deep learning artificial intelligence model is based on a linear stacking model, which involves stacking multiple neural network layers sequentially to form a complete neural network model. In the model, each layer has only one input tensor and one output tensor.

[0110] The underlying algorithm is backpropagation, also known as error backpropagation. It iteratively adjusts the weights and bias parameters in the neural network to make the network's output closer to the actual result. In backpropagation, the values ​​of these parameters are updated by calculating the gradient of the loss function with respect to each parameter in the neural network, causing the loss function to gradually decrease until it reaches its minimum value.

[0111] Models can be constructed by adding different types of layers to create different neural network structures, such as fully connected layers, convolutional layers, pooling layers, and recurrent layers. These different types of layers can be combined to achieve different model architectures and application scenarios.

[0112] The key steps for setting up this artificial intelligence are as follows:

[0113] (1) A large number of photos of the ladle numbers were taken using a process camera, and these photos were manually labeled to form a training set of labels.

[0114] (2) The dataset is divided into a training set and a test set, with the training set accounting for 70% of the total number of photos and the test set accounting for 30% of the total number of photos.

[0115] (3) Organize the training set into a (42, 42, 1) format according to the length, width, and grayscale of the images. The matrix format of the images is a two-dimensional matrix with 42 rows, 42 columns, and 1 layer.

[0116] (4) Use the formula X_train / = 255 # to normalize the pixel values ​​of the training set data to between 0 and 1. This is a matrix division, dividing 255 by each element in the matrix.

[0117] (5) Convert the labels to one-hot encoding to facilitate training the model for multi-class classification tasks. It represents each category as a binary vector, where only one element is 1 and the rest are 0. Specifically, for a classification problem with n categories, each category is represented by a vector of length n, where the i-th element of the vector is 1, indicating that the vector represents the i-th category, and all other elements are 0. This encoding method converts the classification labels into a form that is easier for machine learning algorithms to process, while also improving the accuracy of the classifier.

[0118] (6) Add a convolutional layer with 64 kernels and a kernel size of 3x3 to enable the neural network to better fit the nonlinear data distribution and thus improve the model's performance. In the deep learning of this invention, ReLU (Rectified Linear Unit) is used as one of the activation functions because it has advantages such as simplicity, fast computation speed, and effective prevention of gradient vanishing.

[0119] (7) In order to increase the computation speed of the artificial intelligence model of the present invention, a 2x2 max pooling layer is added. During training, the model’s sensitivity to spatial location can be reduced, and the robustness of the model can be improved.

[0120] (8) To prevent overfitting, two Dropout layers were added during training. These layers randomly set the output of neurons to 0 with a certain probability, thus ensuring that each neuron has a certain probability of being ignored. This forces the network to avoid over-relying on certain neurons during learning, improving the model's generalization ability and reducing the risk of overfitting. The dropout rate of the first Dropout layer in this invention was set to 0.25. The second Dropout layer was placed after the fully connected layer, with a dropout rate set to 0.5.

[0121] Follow the steps above to compile a deep learning artificial intelligence neural network model to identify the ladle number.

[0122] 2. The molten iron ladle is hoisted to the front of the converter. The visual recognition unit at the converter iron-addition position identifies the ladle number and the weight weighed by the crane scale during the iron-addition stage. The ladle number, molten iron weight, and iron-addition time information are then transmitted to the molten iron information integration and transmission unit.

[0123] The iron-feeding position visual recognition unit mainly uses an industrial visual recognition camera to determine the iron-feeding status using the converter iron-feeding logic, and automatically recognizes the numbers marked on the molten iron ladle through an image processing program and a deep learning artificial intelligence model.

[0124] The converter molten iron charging logic refers to identifying the ladle number information during the molten iron charging process. Its sensitive area is a long, narrow rectangle, with zone 1 closer to the converter and zone 2 further away. The molten iron charging process begins when the ladle number appears successively in zones 2 and 1. At this time, a deep learning artificial intelligence model records the ladle number information, time information, and overhead crane weighing information. T兑 (n) represents the weight of the ladle before the iron exchange begins, T 兑 'n' represents the start time of molten iron charging, and 'n' represents the ladle number. The molten iron charging process ends when the ladle number appears successively in zones 1 and 2. At this point, the ladle number information, time information, and overhead crane weighing information are recorded. W T结 (n) represents the final weight of the iron ladle after the iron exchange, T 结 'n' represents the end time of molten iron exchange, and 'n' represents the ladle number.

[0125] The image processing program and deep learning artificial intelligence model are the same as those used for the can-turning position.

[0126] 3. The molten iron information integration and transmission unit mainly uses a computer to analyze and integrate the information provided by the previous units of the system. Utilizing the logic of molten iron composition and temperature, and the logic of molten iron weight, it ultimately transmits the comprehensive information of the composition, temperature, and weight of the molten iron used in the converter to the converter secondary system for converter secondary model calculation.

[0127] The logic for the molten iron composition and temperature is as follows: the time relationship between the molten iron ladle number and the corresponding ladle number in the laboratory is T. 吊 (n) ≤ T 验 (n) ≤ T 兑 At time (n), the chemical composition and temperature corresponding to iron ladle number n are T. 验 (n) Information corresponding to time point T. 吊 (n) represents the time T is spent in the flipping position of the iron bag with bag number n. 兑 (n) represents the time T is spent at the iron exchange point for iron package number n. 验 (n) represents the time it takes for the iron bag with package number n to be analyzed in the laboratory.

[0128] The weight logic for the molten iron is: W 铁 (n) = WT兑 (n)-W T结 (n) 。 Among them W 铁 (n) represents the weight of the molten iron in ladle number n, W T兑 (n) represents the weight of the molten iron before the start of the iron-making process, W. T结 (n) represents the weight of the molten iron after the iron-making process is completed. Example

[0129] Molten iron was transported to the steel plant's ladle transfer station via torpedo ladles, and loaded onto a No. 5 ladle on-site. The molten iron temperature was measured at 1367℃. A sample of the molten iron was sent to the laboratory for testing. During the hoisting process, the ladle number was visually identified as No. 5, and the hoisting time was 14:45. Figure 8 Among them, identifying sensitive areas such as Figure 5 Grayscale processing, such as Figure 6 Binarization processing, such as Figure 7 The iron was hoisted to the converter for ferroalling. Ferrroalling began at 15:00 and ended at 15:03. The identified ladle number was 5, and the identification result was consistent with that of the ladle transfer station. Sensitive areas were identified as follows: Figure 9 The crane weighed 428.5 tons at the start of the molten iron exchange and 163.7 tons at the end. After the exchange was completed, the weight of the molten iron ladle was identified as 264.8 tons. Figure 10 The hot metal composition analysis was completed at 14:55. The final hot metal information for this furnace was integrated and transmitted via the hot metal information integration and transmission unit as follows:

[0130] .

Claims

1. A method for corresponding molten iron information based on visual recognition, characterized in that... Includes the following steps: (1) Molten iron is transported from the torpedo ladle to the ladle-turning station of the steel plant for molten iron turning operation. After the turning operation is completed, the molten iron ladle is driven out by the ground car for temperature measurement and sampling. The ladle number and hoisting time are collected by the visual recognition unit and the recognized molten iron ladle number and hoisting time are transmitted to the molten iron information integration and transmission unit. The molten iron sample is sent to the testing station for testing. The temperature and corresponding ladle number information are transmitted to the laboratory analysis system. After the laboratory composition analysis is completed, the temperature, ladle number and composition information are uploaded to the molten iron information integration and transmission unit. The molten iron ladle pouring visual recognition unit uses an industrial camera to identify the process of hoisting the molten iron ladle out. It automatically identifies the running status of the molten iron ladle through hoisting and return recognition logic, and automatically recognizes the numbers marked on the molten iron ladle through an image processing program and a deep learning artificial intelligence model. The hoisting and return recognition logic is as follows: the sensitive area for recognition is set as a vertically elongated rectangle, which is divided into upper and lower parts from the center. The upper part is zone 1 and the lower part is zone 2. When the ladle number recognition number appears in zone 2 and zone 1 in sequence, it is the hoisting process after the molten iron ladle is turned over. At this time, the molten iron ladle number and hoisting time are recorded. When the ladle number recognition number appears in zone 1 and zone 2 in sequence, it is the empty ladle return process. At this time, no data recording is required. (2) The molten iron ladle is hoisted to the front of the converter furnace. The molten iron ladle number and the weight weighed by the crane scale are identified by the visual recognition unit of the converter iron-pouring position. The molten iron ladle number, molten iron weight and iron-pouring time information are transmitted to the molten iron information integration and transmission unit. The iron-adding position visual recognition unit uses an industrial vision recognition camera and the converter iron-adding logic to determine the iron-adding status. It automatically recognizes the numbers on the molten iron ladle using an image processing program and a deep learning artificial intelligence model. The converter iron-adding logic identifies the ladle number information during the iron-adding process. Its sensitive area is a long, narrow rectangle, with zone 1 closer to the converter and zone 2 further away. When the ladle number appears in zones 2 and 1 successively, the iron-adding process begins. At this time, the deep learning artificial intelligence model records the ladle number information, time information, and overhead crane weighing information. When the ladle number appears in zones 1 and 2 successively, the iron-adding process ends. At this time, the ladle number information, time information, and overhead crane weighing information are recorded. The image processing program and the deep learning artificial intelligence model are the same as those used for the inverted can position; (3) The molten iron information integration and transmission unit analyzes and integrates the information provided by the previous units of the system through the computer, and uses the molten iron composition and temperature logic and molten iron weight logic to finally transmit the comprehensive information of molten iron composition, temperature and weight used in the converter to the converter secondary system for converter secondary model calculation.

2. The method for corresponding molten iron information based on visual recognition according to claim 1, characterized in that: In step (1), the image processing program, while setting the sensitive area, also considers improving computational efficiency and increasing the speed of ladle number recognition. Once it is recognized that there is a ladle in the photo and the ladle has reached the designated position, the ladle photo will be processed as follows: ① Grayscale processing: First, the photo is inverted to grayscale. The value of all pixels in the image is 255 - the current pixel value. ② Binarization processing: Gaussian filtering is performed on all pixels of the image in 3*3 pixel units; ③ The image pixels are eroded in 3*3 pixel units, and the value of the processed pixel is reduced to the minimum value of its neighboring pixels; ④ Perform adaptive filtering on the image. The pixel value of the processed pixel = the average value of adjacent pixels in a 3*3 pixel unit - a constant, where the constant is defined as 10. ⑤ The edge-finding process is used to find the target number; ⑥ Perform contour processing to select the identified digits. Extract the digits from the sensitive area.

3. The method for corresponding molten iron information based on visual recognition according to claim 1, characterized in that: In step (1), the key steps for setting up the deep learning artificial intelligence model are as follows: (a) A large number of photos of ladle numbers were taken using an industrial camera, and these photos were manually labeled to form a training set of labels; (b) The dataset is divided into a training set and a test set, with the training set accounting for 70% of the total number of photos and the test set accounting for 30% of the total number of photos; (c) Organize the training set into a (42, 42, 1) format according to the length, width and grayscale of the images. The matrix format of the images is a two-dimensional matrix with 42 rows, 42 columns and 1 layer. (d) Use the formula X_train / = 255 # to normalize the pixel values ​​of the training set data to between 0 and 1. This is a matrix division, dividing 255 by each element in the matrix. (e) Convert the labels into one-hot encoded form to facilitate the training of the model for multi-class classification tasks; it represents each category as a binary vector, in which only one element is 1 and the rest are 0; specifically, for a classification problem with n categories, each category is represented by a vector of length n, the i-th element of the vector is 1, indicating that the vector represents the i-th category, while the other elements are 0. (f) Establish a sequential model to facilitate the design of deep learning neural network architecture; (g) Add a convolutional layer with 32 kernels, a kernel size of 3x3, and the activation function ReLU. The input data shape is (28, 28, 1). (h) Add another convolutional layer with 64 kernels, a kernel size of 3x3, and the activation function ReLU; (i) Add a 2x2 max pooling layer; (j) Add a dropout layer with a dropout rate of 25%; (k) Flatten the data into 1D data; (l) Add a fully connected layer with 128 nodes and ReLU activation function; (m) Add a dropout layer with a dropout rate of 50%; (n) Add an output layer with 10 nodes and softmax activation function; (o) Compilation model, loss function is cross-entropy, optimizer is Adam, and evaluation metric is accuracy.

4. The method for corresponding molten iron information based on visual recognition according to claim 1, characterized in that: In step (3), the logic for the composition and temperature of molten iron is as follows: the time relationship between the ladle number and the corresponding ladle number in the laboratory is T. 吊 (n) ≤ T 验 (n) ≤ T 兑 At time (n), the chemical composition and temperature corresponding to iron ladle number n are T. 验 (n) Information corresponding to time T, where T 吊 (n) represents the time T is spent in the flipping position of the iron bag with bag number n. 兑 (n) represents the time T is spent at the iron exchange point for iron package number n. 验 (n) represents the time it takes for the iron bag with package number n to be analyzed in the laboratory; The weight logic for molten iron is: W 铁 (n) = W T兑 (n)-W T结 (n), where W 铁 (n) represents the weight of the molten iron in ladle number n, W T兑 (n) represents the weight of the molten iron before the start of the iron-making process, W. T结 (n) represents the weight of the molten iron after the iron-making process is completed.

Images