Converter steelmaking real-time carbon temperature prediction model, prediction method and system based on Vision Transform

Abstract

The invention provides a converter steelmaking real-time carbon temperature prediction model, prediction method and system based on Vision Transform, and belongs to the field of converter steelmaking. In the prediction model, an image set is input through a first interface, an image convolution layer performs feature extraction to form an image block matrix, and classification lexical elements are added through an image embedding layer to generate an embedding matrix; outputting an image static coding classification lexical element matrix through a static feature encoder; the second interface inputs a carbon temperature value of a previous moment of the corresponding image set; weight parameters are given to the carbon temperature values through a historical information embedding module, coding and mapping are carried out, and then a historical embedding matrix is output by aligning dimensions; and the splicing module splices the image static coding classification lexical element matrix and the historical embedding matrix and adds dynamic feature classification lexical elements, the dynamic feature encoder outputs the dynamic coding classification lexical elements, and an MLP layer classification head generates and outputs a carbon temperature prediction value at the current moment. According to the method, the accuracy, the real-time performance and the continuity of a prediction result are improved.

Description

Technical Field

[0001] This invention belongs to the field of converter steelmaking, specifically relating to a real-time carbon temperature prediction model, prediction method and system for converter steelmaking based on Vision Transformer. Background Technology

[0002] Converter steelmaking is a process that uses oxygen blowing and other methods at high temperatures to oxidize and remove impurities (such as carbon, silicon, and manganese) from molten iron, ultimately obtaining steel that meets the required standards. In this process, carbon content and temperature (hereinafter referred to as carbon temperature) directly affect the composition and physical properties of the molten steel, making parameter control crucial. In traditional production, technicians often rely on manual observation of the fire to determine the converter's carbon temperature. This method depends on worker experience and is highly subjective, resulting in inaccurate carbon temperature control. With the application of intelligent monitoring equipment in production, methods for predicting carbon temperature using machine vision have gradually developed. With the development of deep learning technology, more and more visual neural network models are being applied to predict carbon temperature in converter steelmaking, using video footage of the furnace flame combined with visual feature extraction models for prediction. However, due to the irregular changes in flame characteristics (such as brightness and texture) during actual steelmaking and the high requirements for real-time accuracy in prediction, vision-based carbon temperature prediction in converter steelmaking still faces many algorithmic challenges. In particular, for furnace flames under different furnace conditions, the model needs to learn features that are both generalizable and robust in order to accurately predict carbon temperature.

[0003] In existing technologies, most carbon temperature prediction schemes for the end point of converter steelmaking only predict carbon content a few seconds before the endpoint. They cannot dynamically adjust the raw materials based on real-time furnace conditions, resulting in poor process controllability. For example, adding oxygen when the carbon content is too high is not feasible. Furthermore, vision-based methods for predicting carbon temperature at the end point of converter steelmaking mostly rely on traditional feature extraction methods and convolutional neural network models. While some researchers have proposed models for real-time carbon temperature prediction in converter steelmaking, these models still face challenges such as numerous parameters, complex prediction processes, and poor dynamic controllability. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a real-time carbon temperature prediction model, method, and system for converter steelmaking based on the Vision Transformer. A Furnace Transformer model is constructed based on the Vision Transformer model to predict real-time carbon temperature. By extracting the static and dynamic features of the furnace flame, real-time autoregressive prediction of carbon temperature within 3 minutes before the end of the converter steelmaking process is achieved. This increases the controllability of the steelmaking process while ensuring accuracy, thereby improving the level of automation in metallurgical production.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows:

[0006] In a first aspect, embodiments of the present invention provide a real-time carbon temperature prediction model for converter steelmaking based on Vision Transformer. The model includes: a first interface, an image convolutional layer, an image embedding layer, a static feature encoder, a second interface, a historical information embedding module, a stitching module, a dynamic feature encoder, an MLP layer classification head, and an output interface. The first interface is sequentially connected to the image convolutional layer, the image embedding layer, and the static feature encoder. The second interface is connected to the historical information embedding module. The static feature encoder and the historical information embedding module are simultaneously connected to the stitching module. The stitching module is sequentially connected to the dynamic feature encoder, the MLP layer classification head, and the output interface.

[0007] The first interface is used to input a set of images;

[0008] The image convolutional layer is used to extract features from the image set input through the first interface to form an image block matrix;

[0009] The image embedding layer is used to add static feature classification terms to the image block matrix to generate an embedding matrix;

[0010] The static feature encoder is used to encode the embedding matrix and output an image static encoding classification lexical matrix containing static features;

[0011] The second interface is used to input the carbon temperature value, and the carbon temperature value corresponds to the moment before the first frame image in the image set in the first interface;

[0012] The historical information embedding module is used to assign weight parameters to the carbon temperature value input from the second interface and encode it, and then output the historical embedding matrix after mapping and aligning the dimensions.

[0013] The splicing module is used to splice the image static coding classification term matrix and the historical embedding matrix, and add dynamic feature classification terms to obtain the historical output joint vector;

[0014] The dynamic feature encoder is used to encode the historical output joint vector and output dynamically encoded classification words;

[0015] The MLP layer classification head is used to generate the current carbon temperature prediction value based on dynamically encoded classification terms;

[0016] The output interface is used to output the predicted carbon temperature value, and also to use the current predicted carbon temperature value as the historical carbon temperature value for the next moment, and to incorporate it into the carbon temperature value set of the second interface.

[0017] In a preferred embodiment of the present invention, the static feature encoder and the dynamic feature encoder aggregate embedding matrix information through the classification lexical unit cls_token.

[0018] Secondly, embodiments of the present invention also provide a method for real-time carbon temperature prediction in converter steelmaking based on the Vision Transformer model, the method comprising:

[0019] Step S1: Collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production.

[0020] Step S2: Slide a sliding window of width N on the historical flame image sequence. Each sliding step yields a historical image set consisting of N flame images as the first input. Using the time corresponding to the first frame in each historical image set as the segmentation point, extract the carbon temperature values ​​of all images acquired at the historical time before the segmentation point as the second input.

[0021] Step S3: Extract the carbon temperature value at the split point as the output, and construct a data pair [(first input, second input), output]; construct a dataset based on all historical production data with data pairs as the basic unit, and divide it into training set, test set and validation set;

[0022] Step S4: Construct the carbon temperature prediction model as described above based on the Vision Transformer model;

[0023] Step S5: The carbon temperature prediction model is trained, tested, and validated using a training set to obtain a mature carbon temperature prediction model.

[0024] Step S6: Determine the start and end times of the time period to be predicted; obtain the carbon temperature value of the furnace opening to be predicted at the start time as the initial historical carbon temperature value.

[0025] Step S7: Starting from the beginning time, flame images of the furnace opening to be predicted are acquired in real time at predetermined time intervals, excluding the images at the beginning time. A sliding window with a width of N slides over the sequence of flame images acquired in real time. Each sliding step yields a set of predicted images consisting of N frames of flame images.

[0026] Step S8: Input the current set of predicted images into the first interface of the carbon temperature prediction model, and obtain the classification word matrix through the image convolutional layer, image embedding layer and static feature encoder; input the historical carbon temperature values ​​of all prediction times before the current set of predicted images into the second interface. If the current set of predicted images is the first set after the start time, then all historical carbon temperature values ​​correspond to the initial historical carbon temperature values, and obtain the historical embedding matrix through the historical information embedding module.

[0027] Step S9: After concatenating the classification term matrix and the historical embedding matrix, add dynamic feature classification terms to obtain the historical output joint vector, input it into the dynamic feature encoder, and output the dynamic encoded classification term; then generate the carbon temperature prediction value at the current time through the MLP layer classification head and output it; increment the time value by 1, use the current predicted carbon temperature value as the historical carbon temperature value at the next time, and return to step S7 to realize real-time continuous prediction until the end time.

[0028] In a preferred embodiment of the present invention, step S1 collects a sequence of flame images and the carbon temperature value at each corresponding moment within a predetermined time period before tapping.

[0029] In a preferred embodiment of the present invention, the predetermined time period is 3 minutes before steel tapping.

[0030] In a preferred embodiment of the present invention, the training process in step S5 includes:

[0031] The historical image set, which is used as the first input of the data pair, is input into the first interface. The historical image set is processed by the image convolutional layer to extract features and form an image block matrix. Classification terms are added to the matrix in the image embedding layer to form an embedding matrix. The matrix is ​​input into the static feature encoder and outputs an image static encoding classification term matrix containing static features.

[0032] Meanwhile, the carbon temperature value, which is used as the second input in the data pair, is input into the second interface and then into the historical information embedding module. The carbon temperature values ​​of all images acquired in the historical time before the segmentation point are assigned weight parameters and encoded. After mapping, the historical embedding matrix is ​​output with the aligned dimensions.

[0033] The stitching module stitches the image static coding classification word matrix and the historical embedding matrix together, inputs them into the dynamic feature encoder, and finally outputs the dynamic coding classification words. The MLP layer classification head then generates the carbon temperature prediction value at the current moment and outputs it.

[0034] After comparing the output carbon temperature value with the true value, the parameters in the static feature encoder, weight parameters, MLP fusion module, dynamic feature encoder, and MLP classification head are adjusted to complete the training of the model. The model is then tested and validated on the test set and validation set to obtain a mature carbon temperature prediction model.

[0035] In a preferred embodiment of the present invention, the historical information embedding module selects unique thermal encoding when encoding the carbon temperature value input from the second interface.

[0036] In a preferred embodiment of the present invention, the historical information embedding module is implemented using a multilayer perceptron (MLP); through the mapping of the historical information embedding module composed of the MLP, the dimension of the encoding is aligned with the dimension of the static encoded classification lexical units of the input image at the current moment.

[0037] In a preferred embodiment of the present invention, the predetermined time interval mentioned in step S7 is the time interval between each frame of flame images.

[0038] Thirdly, embodiments of the present invention also provide a real-time carbon temperature prediction system for converter steelmaking based on the Vision Transformer model. The system includes: a time determination module, a data acquisition module, an image set acquisition module, a carbon temperature value acquisition module, a training set construction module, a prediction model construction module, a prediction model training module, and a carbon temperature value prediction module; wherein...

[0039] The time determination module is used to determine the start time and end time of the time period to be predicted;

[0040] The data acquisition module is used to collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production.

[0041] The image set acquisition module is used to slide a sliding window of width N on the historical flame image sequence. Each sliding step obtains a historical image set consisting of N frames of flame images as the first input. It is also used to collect flame images of the furnace mouth to be predicted in real time from the starting time at a predetermined time interval, excluding the image at the starting time.

[0042] The carbon temperature value acquisition module is used to extract the carbon temperature values ​​of all historical images acquired before the segmentation point as the second input, taking the time corresponding to the first frame image in each historical image set as the segmentation point; it is also used to acquire the carbon temperature value of the furnace opening to be predicted at the starting time as the initial historical carbon temperature value, and to slide a sliding window with a width of N on the real-time acquired flame image sequence, with each sliding step obtaining a prediction image set consisting of N flame images.

[0043] The training set construction module is used to extract the carbon temperature value at the split point as the output, construct data pairs [(first input, second input), output]; and construct a dataset based on all historical production data with data pairs as primitives, and divide it into training set, test set and validation set.

[0044] The prediction model building module is used to build the carbon temperature prediction model as described above based on the Vision Transformer model;

[0045] The prediction model training module is used to train, test, and validate the carbon temperature prediction model using a training set to obtain a mature carbon temperature prediction model.

[0046] The carbon temperature prediction module is used to predict and output the carbon temperature value at the current moment based on the predicted image set and the corresponding historical initial and historical carbon temperature values; it is also used to use the currently predicted carbon temperature value as the historical carbon temperature value at the next moment when the time value is increased by 1.

[0047] The solutions of the embodiments of the present invention have the following beneficial effects:

[0048] The real-time carbon temperature prediction model, method, and system for converter steelmaking based on Vision Transformer provided in this invention constructs a Furnace Transformer model. It inputs continuous frame images of the furnace flame in a sliding window format, constructs static and dynamic feature encoders, and extracts features from the flame images at multiple scales. During the smelting process, the carbon content and temperature values ​​exhibit a monotonic changing trend. The model's historical outputs are fed back to the input; that is, all output results between the model's first prediction and the previous prediction are jointly mapped and connected to the input of the dynamic encoder. This allows the model to consider historical information, achieving autoregressive prediction with temporal relevance.

[0049] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of the real-time carbon temperature prediction model for converter steelmaking based on Vision Transformer as described in an embodiment of the present invention.

[0052] Figure 2 This is a flowchart of the real-time carbon temperature prediction method for converter steelmaking based on Vision Transformer as described in the embodiments of the present invention;

[0053] Figure 3 This is a schematic diagram illustrating the training principle of the real-time carbon temperature prediction model as described in an embodiment of the present invention.

[0054] Figure 4 This is a schematic diagram of the data input format of the static feature encoder according to an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the data input format of the dynamic feature encoder according to an embodiment of the present invention. Detailed Implementation

[0056] After discovering the aforementioned problems, the inventors of this application conducted a detailed study on existing real-time carbon temperature prediction methods for converter steelmaking. The study found that existing visual models for converter flame carbon temperature prediction are point-to-point, meaning that an image at a single point in time is input into the model, and the model provides a prediction. However, according to the combustion mechanism, the carbon content of the flame in converter steelmaking shows a continuous decreasing trend over time, while the temperature shows a continuous increasing trend. This indicates that past carbon temperature information affects the predicted carbon content and temperature at the next moment. Therefore, existing models ignore important historical information and lack temporal sequence. Furthermore, judging the carbon temperature parameter value from the image of the "fire" relies on the static and dynamic characteristics of the flame. Most existing models extract features from the image, construct feature descriptors (operators), and map them to the final carbon content prediction, failing to fully reflect the static and dynamic characteristics of the "fire" corresponding to carbon temperature.

[0057] It should be noted that the defects in the above-mentioned prior art solutions are all the result of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of the present invention in the following text should be the inventors' contributions to the present invention.

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. It should be noted that, without conflict, the embodiments and features in the embodiments of the present invention can also be combined with each other.

[0059] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the invention, the terms "first," "second," "third," "fourth," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0060] To address the problem of real-time carbon temperature prediction in converter steelmaking, this invention provides a real-time carbon temperature prediction model, method, and system based on VisionTransformer. A FurnaceTransformer model is constructed, and the prediction process employs an autoregressive mechanism, using past observations to predict the current value. Historical output information from the prediction model is fed back to the input, achieving continuous real-time prediction inference. This effectively utilizes historical carbon temperature information and exhibits good interpretability. Simultaneously, the FurnaceTransformer model uses two encoders to extract static and dynamic features in two stages. The static feature encoder generates a static code (cls_token) for each image, while the dynamic feature encoder uses the static code and historical output information as input to generate a dynamic code (cls_token). This aggregates static features and historical output value information, allowing the model to retain historical information for future carbon temperature prediction. Finally, the dynamic code (cls_token) is processed by a classification head to complete real-time carbon temperature prediction. This multi-scale approach fuses features from different dimensions, thereby improving the real-time performance and accuracy of carbon temperature prediction.

[0061] like Figure 1 As shown, the real-time carbon temperature prediction model for converter steelmaking based on Vision Transformer includes: constructing a carbon temperature prediction model based on the Vision Transformer model, wherein the model includes a first interface, an image convolutional layer, an image embedding layer, a static feature encoder, a second interface, a historical information embedding module, a splicing module, a dynamic feature encoder, an MLP layer classification head, and an output interface.

[0062] The first interface is connected in sequence to the image convolutional layer, the image embedding layer, and the static feature encoder;

[0063] The second interface is connected to the historical information embedding module;

[0064] The static feature encoder and the historical information embedding module are both connected to the splicing module;

[0065] The splicing module is connected in sequence to the dynamic feature encoder, the MLP layer classification head, and the output interface.

[0066] The first interface is used to input a set of images;

[0067] The image convolutional layer is used to extract features from the image set input through the first interface to form an image block matrix;

[0068] The image embedding layer is used to add static feature classification terms to the image block matrix to generate an embedding matrix;

[0069] The static feature encoder is used to encode the embedding matrix and output an image static encoding classification lexical matrix containing static features;

[0070] The second interface is used to input the carbon temperature value, and the carbon temperature value corresponds to the moment before the first frame image in the image set in the first interface;

[0071] The historical information embedding module is used to assign weight parameters to the carbon temperature value input from the second interface and encode it, and then output the historical embedding matrix after mapping and aligning the dimensions.

[0072] The splicing module is used to splice the image static coding classification term matrix and the historical embedding matrix, and add dynamic feature classification terms to obtain the historical output joint vector;

[0073] The dynamic feature encoder is used to encode the historical output joint vector and output dynamically encoded classification words;

[0074] The MLP layer classification head is used to generate the current carbon temperature prediction value based on dynamically encoded classification terms;

[0075] The output interface is used to output the predicted carbon temperature value, and also to use the current predicted carbon temperature value as the historical carbon temperature value for the next moment, and to incorporate it into the carbon temperature value set of the second interface.

[0076] The static and dynamic feature encoders share the same self-attention mechanism encoding principle as the Vision Transformer model, aggregating embedded matrix information through the categorical token cls_token. For example, the categorical token in the dynamic encoder aggregates static feature change information from multiple current input images as well as historical output carbon content value change information, thereby learning temporally sequential dynamic features.

[0077] like Figure 2 As shown, the real-time carbon temperature prediction method for converter steelmaking based on the Vision Transformer model includes the following steps:

[0078] Step S1: Collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production.

[0079] In this step, the carbon temperature value corresponds to the flame image at the current moment in the historical production process. During converter production, what generally needs to be detected or predicted is the carbon content and temperature value within a predetermined time period before tapping, for example, three minutes before tapping.

[0080] Step S2: Slide a sliding window of width N across the historical flame image sequence. Each sliding step yields a historical image set consisting of N flame images as the first input. Using the time corresponding to the first frame in each historical image set as the segmentation point, extract the carbon temperature values ​​of all images acquired at historical times before the segmentation point as the second input.

[0081] Step S3: Extract the carbon temperature value at the split point as the output, and construct a data pair [(first input, second input), output]; construct a dataset based on all historical production data with data pairs as primitives, and divide it into training set, test set and validation set.

[0082] Step S4: Construct the carbon temperature prediction model as described above based on the Vision Transformer model. The model includes a first interface, an image convolutional layer, an image embedding layer, a static feature encoder, a second interface, a historical information embedding module, a stitching module, a dynamic feature encoder, an MLP layer classification head, and an output interface.

[0083] The carbon temperature prediction model mentioned in this step is the real-time carbon temperature prediction model for converter steelmaking based on VisionTransformer provided in the embodiments of this application, which will not be described again here.

[0084] Step S5: The carbon temperature prediction model is trained, tested, and validated using a training set to obtain a mature carbon temperature prediction model.

[0085] like Figure 3 As shown, during training, the historical image set, which is the first input of the data pair, is input into the first interface. The historical image set is processed by the image convolutional layer to extract features and form an image block matrix. In the image embedding layer, classification terms are added to the matrix to form an embedding matrix, which is input into the static feature encoder and outputs an image static encoding classification term matrix containing static features.

[0086] like Figure 4 As shown, the input data of the static feature encoder includes two dimensions: feature dimension and embedding vector dimension. The feature dimension comes from the image patch matrix extracted from each frame. For example, the matrix includes d1 to d2 in the column direction. n If there are n features, then the feature dimension is n; the embedding vector dimension includes the length of the image patch matrix in the embedding direction and the length of the added classification terms. When adding classification terms, each frame in the image set is added separately, and the aggregated terms together form the embedding matrix.

[0087] Simultaneously, the carbon temperature value, used as the second input in the data pair, is input to the second interface, and then to the historical information embedding module. This module assigns weight parameters to the carbon temperature values ​​from all historical image acquisition times prior to the segmentation point and encodes them. After mapping, the dimensions are aligned, and the historical embedding matrix is ​​output. Preferably, one-hot encoding is used to ensure that one carbon content parameter corresponds to one unit vector. The historical information embedding module is implemented using a multilayer perceptron (MLP). Through the mapping of the historical information embedding module composed of the MLP, the encoded dimension is aligned with the dimension of the statically encoded classification terms of the input image at the current moment.

[0088] The concatenation module concatenates the image static encoding classification term matrix and the historical embedding matrix, inputs them into the dynamic feature encoder, and finally outputs the dynamic encoding classification term. The MLP layer classification head generates the carbon temperature prediction value at the current moment, thus effectively utilizing the historical carbon temperature information and making the model temporal.

[0089] like Figure 5 As shown, the input data format of the dynamic feature encoder is similar to that of the static feature encoder, including feature dimension and concatenation vector dimension. The concatenation vector dimension is the sum of the length of the historical embedding matrix, the length of the static classification lexical matrix, and the length of the added dynamic feature classification lexical.

[0090] After comparing the output carbon temperature value with the true value, the parameters in the static feature encoder, weight parameters, MLP fusion module, dynamic feature encoder, and MLP classification head are adjusted to complete the training of the model. The model is then tested and validated on the test set and validation set to obtain a mature carbon temperature prediction model.

[0091] Step S6: Determine the start and end times of the time period to be predicted; obtain the carbon temperature value of the furnace opening to be predicted at the start time as the initial historical carbon temperature value.

[0092] Step S7: Starting from the beginning time, flame images of the furnace opening to be predicted are acquired in real time at predetermined time intervals, excluding the images at the beginning time. A sliding window with a width of N slides across the sequence of flame images acquired in real time. Each sliding step yields a set of predicted images consisting of N frames of flame images.

[0093] Step S8: Input the current set of predicted images into the first interface of the carbon temperature prediction model, and obtain the classification word matrix through the image convolutional layer, image embedding layer and static feature encoder; input the historical carbon temperature values ​​of all prediction times before the current set of predicted images into the second interface. If the current predicted image is the first group after the start time, then all historical carbon temperature values ​​correspond to the initial historical carbon temperature values, and obtain the historical embedding matrix through the historical information embedding module.

[0094] Step S9: After concatenating the classification term matrix and the historical embedding matrix, add dynamic feature classification terms to obtain the historical output joint vector, input it into the dynamic feature encoder, and output the dynamic encoded classification term; then generate the carbon temperature prediction value at the current time through the MLP layer classification head and output it; increment the time value by 1, use the current predicted carbon temperature value as the historical carbon temperature value at the next time, and return to step S7 to realize real-time continuous prediction until the end time.

[0095] Based on the same idea, this invention also provides a real-time carbon temperature prediction system for converter steelmaking based on the Vision Transformer model. The system includes: a time determination module, a data acquisition module, an image set acquisition module, a carbon temperature value acquisition module, a training set construction module, a prediction model construction module, a prediction model training module, and a carbon temperature value prediction module; wherein...

[0096] The time determination module is used to determine the start and end times of the time period to be predicted; it can also be used to determine the time period of historical production data.

[0097] The data acquisition module is used to collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production.

[0098] The image set acquisition module is used to slide a sliding window of width N on the historical flame image sequence. Each sliding step obtains a historical image set consisting of N frames of flame images as the first input. It is also used to collect flame images of the furnace mouth to be predicted in real time from the starting time at a predetermined time interval, excluding the image at the starting time.

[0099] The carbon temperature value acquisition module is used to extract the carbon temperature values ​​of all historical images acquired before the segmentation point as the second input, taking the time corresponding to the first frame image in each historical image set as the segmentation point; it is also used to acquire the carbon temperature value of the furnace opening to be predicted at the starting time as the initial historical carbon temperature value, and to slide a sliding window with a width of N on the real-time acquired flame image sequence, with each sliding step obtaining a prediction image set consisting of N flame images.

[0100] The training set construction module is used to extract the carbon temperature value at the split point as the output, construct data pairs [(first input, second input), output]; and construct a dataset based on all historical production data with data pairs as primitives, and divide it into training set, test set and validation set.

[0101] The prediction model building module is used to build the carbon temperature prediction model as described above based on the Vision Transformer model;

[0102] The prediction model training module is used to train, test, and validate the carbon temperature prediction model using a training set to obtain a mature carbon temperature prediction model.

[0103] The carbon temperature prediction module is used to predict and output the carbon temperature value at the current moment based on the predicted image set and the corresponding historical initial and historical carbon temperature values; it is also used to use the currently predicted carbon temperature value as the historical carbon temperature value at the next moment when the time value is increased by 1.

[0104] In this embodiment, each module is implemented using a processor, with additional memory added as needed for storage. The processor can be, but is not limited to, a microprocessor (MPU), a central processing unit (CPU), a network processor (NP), a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), other programmable logic devices, discrete gates, transistor logic devices, discrete hardware components, etc. The memory can include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory can also be at least one storage device located remotely from the aforementioned processor.

[0105] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0106] It should also be noted that the real-time carbon temperature prediction system for converter steelmaking based on the Transformer model described in this embodiment corresponds to the real-time carbon temperature prediction method for converter steelmaking based on the Transformer model. The description and limitations of the method also apply to the system, and will not be repeated here.

[0107] Therefore, the real-time carbon temperature prediction model, prediction method and system for converter steelmaking based on the Vision Transformer model provided in this embodiment of the invention constructs a Furnace Transformer model, inputs continuous frame images of the furnace flame video in a sliding window format, constructs static feature and dynamic feature encoders, and extracts features from the flame images at multiple scales. During the smelting process, the carbon content and temperature values ​​show a monotonic changing trend. The historical output of the model is fed back to the input, that is, all the output results between the first prediction of the model and the previous prediction are jointly mapped and connected to the input of the dynamic encoder, so that the model can pay attention to historical information and achieve autoregressive prediction with temporal characteristics.

[0108] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed, and is not intended to limit the scope of the claimed invention, but merely to illustrate preferred embodiments of the invention. Those skilled in the art should understand that the scope of the invention is not limited to the specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A real-time carbon temperature prediction model for converter steelmaking based on Vision Transformer, characterized in that, The model includes: a first interface, an image convolutional layer, an image embedding layer, a static feature encoder, a second interface, a historical information embedding module, a stitching module, a dynamic feature encoder, an MLP layer classification head, and an output interface; the first interface is connected to the image convolutional layer, the image embedding layer, and the static feature encoder in sequence; the second interface is connected to the historical information embedding module; the static feature encoder and the historical information embedding module are both connected to the stitching module; the stitching module is connected to the dynamic feature encoder, the MLP layer classification head, and the output interface in sequence. The first interface is used to input a set of images; The image convolutional layer is used to extract features from the image set input through the first interface to form an image block matrix; The image embedding layer is used to add static feature classification terms to the image block matrix to generate an embedding matrix; The static feature encoder is used to encode the embedding matrix and output an image static encoding classification lexical matrix containing static features; The second interface is used to input the carbon temperature value, and the carbon temperature value corresponds to the moment before the first frame image in the image set in the first interface; The historical information embedding module is used to assign weight parameters to the carbon temperature value input from the second interface and encode it, and then output the historical embedding matrix after mapping and aligning the dimensions. The splicing module is used to splice the image static coding classification term matrix and the historical embedding matrix, and add dynamic feature classification terms to obtain the historical output joint vector; The dynamic feature encoder is used to encode the historical output joint vector and output dynamically encoded classification words; The MLP layer classification head is used to generate the current carbon temperature prediction value based on dynamically encoded classification terms; The output interface is used to output the predicted carbon temperature value, and also to use the current predicted carbon temperature value as the historical carbon temperature value for the next moment, and to incorporate it into the carbon temperature value set of the second interface.

2. The model according to claim 1, characterized in that, The static feature encoder and the dynamic feature encoder aggregate the embedding matrix information through the classification unit cls_token.

3. A method for real-time carbon temperature prediction in converter steelmaking based on the Vision Transformer model, characterized in that, The method includes: Step S1: Collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production. Step S2: Slide a sliding window of width N on the historical flame image sequence. Each sliding step yields a historical image set consisting of N flame images as the first input. Using the time corresponding to the first frame in each historical image set as the segmentation point, extract the carbon temperature values ​​of all images acquired at the historical time before the segmentation point as the second input. Step S3: Extract the carbon temperature value at the split point as the output, and construct a data pair [(first input, second input), output]; construct a dataset based on all historical production data with data pairs as the basic unit, and divide it into training set, test set and validation set; Step S4: Construct the carbon temperature prediction model as described in claim 1 or 2 based on the Vision Transformer model; Step S5: The carbon temperature prediction model is trained, tested, and validated using a training set to obtain a mature carbon temperature prediction model. Step S6: Determine the start and end times of the time period to be predicted; obtain the carbon temperature value of the furnace opening to be predicted at the start time as the initial historical carbon temperature value. Step S7: Starting from the beginning time, flame images of the furnace opening to be predicted are acquired in real time at predetermined time intervals, excluding the images at the beginning time. A sliding window with a width of N slides over the sequence of flame images acquired in real time. Each sliding step yields a set of predicted images consisting of N frames of flame images. Step S8: Input the current set of predicted images into the first interface of the carbon temperature prediction model, and obtain the classification word matrix through the image convolutional layer, image embedding layer and static feature encoder; input the historical carbon temperature values ​​of all prediction times before the current set of predicted images into the second interface. If the current set of predicted images is the first set after the start time, then all historical carbon temperature values ​​correspond to the initial historical carbon temperature values, and obtain the historical embedding matrix through the historical information embedding module. Step S9: After concatenating the classification term matrix and the historical embedding matrix, add dynamic feature classification terms to obtain the historical output joint vector, input it into the dynamic feature encoder, and output the dynamic encoded classification term; then generate the carbon temperature prediction value at the current time through the MLP layer classification head and output it; increment the time value by 1, use the current predicted carbon temperature value as the historical carbon temperature value at the next time, and return to step S7 to realize real-time continuous prediction until the end time.

4. The method according to claim 3, characterized in that, Step S1 collects a sequence of flame images and the carbon temperature value at each corresponding moment within a predetermined time period before tapping.

5. The method according to claim 4, characterized in that, The predetermined time period is 3 minutes before steel tapping.

6. The method according to claim 3, characterized in that, The training process in step S5 includes: The historical image set, which is used as the first input of the data pair, is input into the first interface. The historical image set is processed by the image convolutional layer to extract features and form an image block matrix. Classification terms are added to the matrix in the image embedding layer to form an embedding matrix. The matrix is ​​input into the static feature encoder and outputs an image static encoding classification term matrix containing static features. Meanwhile, the carbon temperature value, which is used as the second input in the data pair, is input into the second interface and then into the historical information embedding module. The carbon temperature values ​​of all images acquired in the historical time before the segmentation point are assigned weight parameters and encoded. After mapping, the historical embedding matrix is ​​output with the aligned dimensions. The stitching module stitches the image static coding classification word matrix and the historical embedding matrix together, inputs them into the dynamic feature encoder, and finally outputs the dynamic coding classification words. The MLP layer classification head then generates the carbon temperature prediction value at the current moment and outputs it. After comparing the output carbon temperature value with the true value, the parameters in the static feature encoder, weight parameters, MLP fusion module, dynamic feature encoder, and MLP classification head are adjusted to complete the training of the model. The model is then tested and validated on the test set and validation set to obtain a mature carbon temperature prediction model.

7. The method according to claim 6, characterized in that, The historical information embedding module selects unique thermal encoding when encoding the carbon temperature value input from the second interface.

8. The method according to claim 3, characterized in that, The historical information embedding module is implemented using a multilayer perceptron (MLP). Through the mapping of the historical information embedding module composed of the MLP, the dimension of the encoding is aligned with the dimension of the static encoded classification words of the input image at the current moment.

9. The method according to claim 3, characterized in that, The predetermined time interval mentioned in step S7 is the time interval between each frame of flame images.

10. A real-time carbon temperature prediction system for converter steelmaking based on the Vision Transformer model, characterized in that, The system includes: a time determination module, a data acquisition module, an image set acquisition module, a carbon temperature value acquisition module, a training set construction module, a prediction model construction module, a prediction model training module, and a carbon temperature value prediction module; wherein, The time determination module is used to determine the start time and end time of the time period to be predicted; The data acquisition module is used to collect the flame image sequence and the carbon temperature value at each corresponding moment in the current converter's historical production. The image set acquisition module is used to slide a sliding window of width N on the historical flame image sequence. Each sliding step obtains a historical image set consisting of N frames of flame images as the first input. It is also used to collect flame images of the furnace mouth to be predicted in real time from the starting time at a predetermined time interval, excluding the image at the starting time. The carbon temperature value acquisition module is used to extract the carbon temperature values ​​of all historical images acquired before the segmentation point as the second input, taking the time corresponding to the first frame image in each historical image set as the segmentation point; it is also used to acquire the carbon temperature value of the furnace opening to be predicted at the starting time as the initial historical carbon temperature value, and to slide a sliding window with a width of N on the real-time acquired flame image sequence, with each sliding step obtaining a prediction image set consisting of N flame images. The training set construction module is used to extract the carbon temperature value at the split point as the output, construct data pairs [(first input, second input), output]; and construct a dataset based on all historical production data with data pairs as primitives, and divide it into training set, test set and validation set. The prediction model building module is used to build the carbon temperature prediction model as described in claim 1 or 2 based on the Vision Transformer model; The prediction model training module is used to train, test, and validate the carbon temperature prediction model using a training set to obtain a mature carbon temperature prediction model. The carbon temperature prediction module is used to predict and output the carbon temperature value at the current moment based on the predicted image set and the corresponding historical initial and historical carbon temperature values; it is also used to use the currently predicted carbon temperature value as the historical carbon temperature value at the next moment when the time value is increased by 1.

Images