Model and method for detecting temperature in converter in later stage of converter steelmaking

Abstract

The invention discloses a model and method for detecting the temperature in a converter in the later stage of converter steelmaking, and relates to the technical field of temperature detection.The model comprises a feature extractor module which is formed by connecting a convolution and pooling layer, four block layers and an average pooling layer front and back and is used for carrying out automatic feature extraction on a furnace mouth flame image; the improved ECA module is included in each block layer and is used for carrying out weighted calculation on different channel weights by carrying out 1D convolution operation on channel dimensions; the regression device module is used for calculating the feature data calculated by the feature extractor module to obtain a normalized dimensionless detection value; during pre-training, the training learning module independently trains the feature extractor module, during formal training, the training learning module simultaneously trains the feature extractor module and the regression device module, and parameter updating is carried out through error back propagation; and the output module performs reverse normalization operation on the dimensionless detection value output by the regression device module to obtain a temperature detection value with a temperature dimension.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature detection, and in particular to a model and method for detecting the temperature in a converter furnace in the later stage of steelmaking. Background Art

[0002] As a heavy industrial country, the steel industry has always been an important part of my country's national economy. In recent years, my country's steel production has always occupied the largest share in the world. In recent years, the country has put forward transformation requirements for the development of the steel industry. While improving the process level and independent innovation capabilities, energy conservation and consumption reduction must be achieved. The metallurgical industry is not only an important part of the steel industry, but its process level is also an important indicator of a country's industrialization level, and steelmaking is an important process in the metallurgical industry.

[0003] At present, the converter steelmaking method is the main steelmaking method used in most countries, including my country, due to its advantages of high production efficiency, fast speed and low investment. In the converter steelmaking process, it is very important to pay attention to the situation in the molten pool in real time. The situation in the furnace not only affects the control operation of the converter at the current moment, but also has a great relationship with the final steel quality. At present, domestic factories still mainly use manual methods, and workers with rich steelmaking experience observe the flame and the characteristics of the molten steel at the furnace mouth to make judgments and adopt corresponding control strategies. The furnace temperature value in the later stage of converter steelmaking is an important reference for steel workers. The accurate grasp of the furnace temperature value will greatly affect the quality of steel. However, this experience-based judgment method places extremely high demands on steel workers. At the same time, due to the need for real-time observation, the judgment results of workers are easily affected by subjective emotions and other factors. The accuracy of judgment based on experience is not high, and real-time observation of the high-brightness furnace mouth is also easy to affect the health of workers, and the consumption of manpower and material resources is large.

[0004] The existing methods for detecting the temperature in the late stage of converter steelmaking are divided into two types: hard measurement and soft measurement. Dong Kai et al. used a temperature detection system to realize the hard measurement of the temperature in the converter furnace in the Chinese invention patent application "A method for controlling the end point of converter steelmaking based on dynamic mixed injection of CO2" (application number: CN202211529994.7). Pan Jun et al. used a direct sampling temperature measurement method to obtain the furnace temperature in the Chinese invention patent application "A method for compensating the end point temperature of converter blowing" (application number: CN201910358182.2). Due to the high temperature in the furnace in the late stage of converter steelmaking, the hard measurement method causes serious loss of the measuring equipment, which increases the production cost. In addition, the delay of hard measurement is large and the real-time performance of the measurement is poor. Most soft measurement methods build models based on production data or manually design image features to build models. Shen Weiming et al. used a network model based on radial basis function in the Chinese invention patent application "A method and device for optimizing converter steelmaking operation based on NSGA-II" (application number: CN202410529431.0) to predict the temperature in the late stage of converter steelmaking. Liu Hui et al. used an ensemble learning method in the Chinese invention patent application "A soft measurement method for carbon temperature at the end of converter steelmaking based on ensemble learning" (application number: CN202310238762.4) to realize the soft measurement of the temperature in the late stage of converter steelmaking. However, the method of building a model based on production data requires a large amount of production data and has high requirements on the equipment of the steel plant. Cheng Xiangming et al. used a method based on traditional image features in the Chinese invention patent application "A prediction system and method for judging the end temperature of converter steelmaking based on flame images" (application number: CN202110944596.0) to extract the brightness value, rotation-invariant LBP eigenvalue and HOG eigenvalue of the flame image for temperature prediction, but the calculation of traditional image features is often complicated and has poor generalization performance.

[0005] In recent years, with the rapid development of deep learning, convolutional neural networks have been widely used in the field of images. Convolutional neural networks avoid the trouble of manually designing features, automatically extract features through convolution kernels, calculate network losses through loss functions, and update model parameters through stochastic gradient descent.

[0006] Therefore, technicians in this field are committed to developing a model and method for detecting the temperature in the later stage of converter steelmaking based on convolutional neural networks to reduce detection costs and improve detection speed. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to realize an accurate and efficient temperature detection method to ensure the accuracy and robustness of the temperature detection in the furnace in the later stage of converter steelmaking and improve the quality of steel output.

[0008] To achieve the above object, the present invention provides a temperature detection model in a converter steelmaking late stage furnace, comprising: The feature extractor module is a multi-layer structure based on a convolutional neural network, which is composed of a convolution and pooling layer, four block layers, and an average pooling layer connected in front and behind, and automatically extracts features from the furnace mouth flame image to obtain feature data; An improved ECA module, which is included in each block layer of the feature extractor module, performs a 1D convolution operation on the channel dimension to perform weighted calculation on different channel weights, thereby improving the validity of the feature data; A regressor module, the regressor module is connected to the feature extractor module, is a double-layer structure based on a fully connected neural network, calculates the feature data calculated by the feature extractor module to obtain a normalized dimensionless detection value; A training learning module, wherein the training learning module connects the feature extractor module and the regressor module; during pre-training, the training learning module trains the feature extractor module alone, and during formal training, the training learning module trains the feature extractor module and the regressor module at the same time, and performs parameter update through error back propagation; An output module is connected to the regressor module, and performs a denormalization operation on the dimensionless detection value output by the regressor module to obtain a temperature detection value with a temperature dimension. The smaller the error between the temperature detection value and the true value, the more accurate the temperature detection of the model is.

[0009] Furthermore, the parameters of the average pooling layer in the feature extractor module are adjusted according to the size characteristics of the input furnace flame image, and the formula is as follows: in, represents the height of the furnace flame image, Indicates the width of the furnace mouth flame image; For the pooling operation class, is an instantiated pooling operation class object. is the feature data before the pooling operation, Realize The pooling operation, is the feature data after the pooling operation, that is, the output of the feature extractor module.

[0010] Furthermore, the improved ECA module performs weighted calculation on different channels, and the formula is as follows: in, The characteristic data before the ECA module is improved. The feature data after passing the improved ECA module; The internal calculation process of the improved ECA module is as follows: in, For the pooling operation class, The function compresses the last dimension of the feature data. The function swaps the last two dimensions of the feature data; The function performs a 1D convolution operation on the last dimension of the feature data to achieve cross-channel information exchange. The function adds a dimension after the last dimension of the feature data. The function is the activation function, The function expands the feature data to the specified dimension; During the entire calculation process, The feature dimension changes are as follows: [B,C,H,W]->[B,C,1,1]->[B,C,1]->[B,1,C]->[B,1,C]->[B,C,1]->[B,C,1,1]->[B,C,H,W] Finally, the calculated attention Weight and multiply, thereby enhancing important features and suppressing unimportant features.

[0011] Furthermore, during pre-training, the pre-training method is self-supervision and the global mutual information loss function is used for optimization, wherein the calculation formula of the global mutual information loss function is as follows: in, is the ReLU function used to approximate the gradient, is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; The objective function of the feature extractor module in the pre-training phase is as follows: pass Calculate the loss value, calculate and update the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module. .

[0012] Furthermore, during the formal training, the training of the regressor module and the training of the feature extractor module are performed simultaneously, and the loss function formula during the training process is as follows: in, Indicates the true value of temperature, represents the temperature detection value output by the regressor module, for The segmentation points of the function; In the late stage of converter steelmaking during formal training, the objective function of the furnace temperature detection model is as follows: pass The function calculates the loss value and calculates and updates the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm. The network parameters of the regressor module , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module. The network parameters of the regressor module .

[0013] The present invention also provides a method for detecting the temperature in a converter steelmaking furnace in the later stage, the method comprising the following steps: Step 1: Collect several furnace mouth flame images in the late stage of converter steelmaking to construct a data set; Step 2: preprocessing the data set by normalization and data enhancement to obtain enhanced data; Step 3, define the temperature detection model in the late stage of converter steelmaking, including determining the input and output of the model, determining the feature extractor module, improved ECA module, regressor module, loss function and objective function in the model; Step 4, dividing the data set into a training set, a validation set and a test set, using the training set to train the converter steelmaking late furnace temperature detection model, and using the validation set to verify the converter steelmaking late furnace temperature detection model; wherein 80% is the training set, 10% is the validation set, and 10% is the test set; Step 5: input the test set into the converter steelmaking late furnace temperature detection model, and output the prediction result of the furnace temperature detection.

[0014] Furthermore, the step 1 includes the following sub-steps: Step 1.1, record the furnace mouth flame video of the converter steelmaking in the later stage through an industrial camera; Step 1.2, select video clips with less image interference and good imaging quality to extract frames, and obtain several furnace mouth flame images; Step 1.3, cropping the furnace mouth flame image, removing invalid areas, and retaining valid areas; Step 1.4, the cropped furnace mouth flame image is matched with the temperature inside the furnace to construct the data set.

[0015] Furthermore, the step 2 includes the following sub-steps: Step 2.1, perform data normalization, the furnace mouth flame image is RGB color space, the color value range is 0 to 255, and the color value range of the training set and the test set is mapped to the (0, 1) interval; Step 2.2, perform data enhancement, including randomly flipping and rotating the furnace mouth flame image horizontally, wherein the probability of random horizontal flipping p=0.5, and the angle range of random rotation is set to (-15°, 15°).

[0016] Furthermore, the 1*1 convolution kernel size of the improved ECA module in step 3 is set to 3, that is, three similar channels are allowed to exchange information; the loss function of the feature extractor module in the pre-training stage is the maximum mutual information loss, and the loss function of the feature extractor module and the regressor module in the joint training stage is ; In the pre-training stage, the objective function of the feature extractor module is as follows: in, and represents the intermediate feature data of the feature extractor module during the calculation process, and Indicates that after Layer and Layer Network Layer, It is an activation function used to approximate the gradient. Its shape is similar to the RELU function but smoother. is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; In the formal training phase, the objective function of the feature extractor module and the regressor module for joint training is as follows: .

[0017] Furthermore, step 4 includes the following sub-steps: Step 4.1, divide the data set into a training set, a validation set and a test set, where 80% is the training set, 10% is the validation set, and 10% is the test set; Step 4.2, pre-training the feature extractor module, including 20 rounds of preheating and 300 rounds of formal training, wherein the batch size is 40, the adam optimizer is used, the initial learning rate of the 20 rounds of preheating is 0.0001, the initial learning rate of the 300 rounds of formal training is 0.001, the cosine annealing learning rate is used, and the loss function is the global mutual information loss; Step 4.3, the feature extractor module and the regressor module are jointly trained, including 20 rounds of preheating and 300 rounds of formal training, wherein the batch size is 40, the adam optimizer is used, the initial learning rate of the 20 rounds of preheating is 0.0001, the initial learning rate of the 300 rounds of formal training is 0.002, the cosine annealing learning rate is used, and the loss function is Huberloss loss; Step 4.4, use evaluation indicators to evaluate the converter steelmaking later furnace temperature detection model, the evaluation indicators include accuracy, mean square error, root mean square error and mean absolute percentage error, measure the average difference between the output value of the converter steelmaking later furnace temperature detection model and the true value, the difference between the model output value and the true value of the temperature is within 15 ° C. The formulas for accuracy, mean square error, root mean square error and mean absolute percentage error are as follows: .

[0018] The present invention provides a converter steelmaking late furnace temperature detection model and method having at least the following technical effects: Traditional hard measurement methods cause serious wear and tear on measurement equipment, increasing production costs; and hard measurement has a large delay and poor real-time measurement. The method of building a model based on production data requires a large amount of production data and has high requirements for steel mill equipment; the method based on manually designed image features not only has complex feature calculations, but also has poor generalization performance. In the technical solution provided by the present invention, a convolutional neural network is applied to the furnace temperature detection model in the later stage of converter steelmaking. Compared with the hard measurement method, it reduces the use of measurement consumables and reduces production costs; compared with the method of building a model based on production data, it greatly reduces the equipment requirements for steel mills; compared with the method based on manually designed image features, it eliminates the tediousness of designing features and improves the speed of feature extraction.

[0019] In order to improve the feature extractor module's ability to understand the furnace mouth flame video, the feature extractor module is pre-trained based on the maximum mutual information theory; an improved ECA module is added to perform weighted summation on different channels of feature data; the size of the last layer of the feature extractor module, i.e., the feature extraction layer, is modified according to actual conditions, so that the feature extractor module can extract the most effective flame image features. In general, the technical solution provided by the present invention has low deployment and use costs, high accuracy and real-time performance of temperature detection, and provides a simple and effective solution for online real-time furnace temperature detection of small and medium-sized converters.

[0020] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a structural schematic diagram of a temperature detection model of a preferred embodiment of the present invention; Figure 2It is a flow chart of a temperature detection method of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0023] Based on the research and imitation of the human visual cortex, the embodiment of the present invention designs a convolutional neural network model for the detection of the temperature in the late stage of converter steelmaking, and applies the convolutional neural network to the detection of the temperature in the late stage of converter steelmaking. The embodiment of the present invention collects the flame image of the converter mouth from the actual production site and performs data processing. First, the feature extractor module is pre-trained based on the maximum mutual information theory, and then the feature extractor module and the regressor module are trained together. The feature extractor module includes an improved ECA module, and the function of the improved ECA module is to modify the weights of different channels by channel convolution; at the same time, according to the characteristics of the flame image, the parameters of the last layer of the feature extractor, that is, the average pooling layer, are modified to (1,4). By taking the above operations, the validity of the data features obtained by the feature extractor module is improved. The regressor calculates according to the feature data provided by the feature extractor and outputs the normalized dimensionless temperature prediction value. Finally, the output module performs a denormalization operation to obtain a dimensional temperature prediction value.

[0024] Example 1 like Figure 1 As shown, an embodiment of the present invention provides a temperature detection model in a converter steelmaking late stage furnace, comprising: The main part of the feature extractor module is composed of four blocks with similar structures stacked together to simplify the network structure and scalability. The feature extractor module is a multi-layer structure based on a convolutional neural network, consisting of a convolution and pooling layer, four block layers, and an average pooling layer connected front and back. It automatically extracts features from the furnace flame image to obtain feature data; Improved ECA module: The improved ECA module is included in each block layer of the feature extractor module. The purpose is to improve the performance of the feature extractor module. By performing 1D convolution operation on the channel dimension, different channel weights are weighted and calculated to improve the validity of feature data. The regressor module is connected to the feature extractor module and is a double-layer structure based on a fully connected neural network. The feature data calculated by the feature extractor module is calculated to obtain a normalized dimensionless detection value. A training learning module, which connects the feature extractor module and the regressor module. During pre-training, the training learning module trains the feature extractor module alone. During formal training, the training learning module trains the feature extractor module and the regressor module at the same time, and updates the parameters through error back propagation. The output module is connected to the regressor module, and the dimensionless detection value output by the regressor module is denormalized to obtain a temperature detection value with a temperature dimension. The smaller the error between the temperature detection value and the true value, the more accurate the temperature detection of the model is.

[0025] In particular, the formula for the feature extractor module is: in, is the input furnace flame image, are the feature extractor parameters, The feature data output by the feature extractor module.

[0026] Example 2 Based on Example 1, the parameters of the average pooling layer in the feature extractor module are adjusted according to the size characteristics of the input furnace flame image, and the formula is as follows: in, Indicates the height of the furnace flame image, Indicates the width of the flame image at the furnace mouth. Since the furnace mouth of the converter is relatively long and narrow, More reasonable; For the pooling operation class, is an instantiated pooling operation class object. is the feature data before the pooling operation, Realize The pooling operation, is the feature data after the pooling operation, that is, the output of the feature extractor module.

[0027] Example 3 On the basis of Example 2, the improved ECA module performs weighted calculation on different channels, and the formula is as follows: in, In order to improve the feature data before the ECA module, It is the characteristic data after improving the ECA module; The internal calculation process of the improved ECA module is as follows: in, For the pooling operation class, The function compresses the last dimension of the feature data. The function swaps the last two dimensions of the feature data; The function performs a 1D convolution operation on the last dimension of the feature data to achieve cross-channel information exchange. The function adds a dimension after the last dimension of the feature data. The function is the activation function, The function expands the feature data to the specified dimension; During the entire calculation process, The feature dimension changes are as follows: [B,C,H,W]->[B,C,1,1]->[B,C,1]->[B,1,C]->[B,1,C]->[B,C,1]->[B,C,1,1]->[B,C,H,W] Finally, the calculated attention Weight and multiply, thereby enhancing important features and suppressing unimportant features.

[0028] In particular, The function is Function improvements, relief Function derivation for complex problems to improve calculation speed, Functions and The function formulas are: You can see The function is a variant of the RELU function and is easy to calculate.

[0029] Example 4 Based on Example 3, the regressor module is a double-layer structure based on a fully connected neural network, and the formula is: in, It is the output feature data of the feature extractor module and also the input data of the regressor module. are the network parameters of the regressor module, is the output of the regressor module, i.e. the normalized predicted temperature.

[0030] During pre-training, the update formula of the feature extractor is based on the maximum mutual information theory. In order to improve the performance of the model, it is necessary to use the furnace flame image to pre-train the feature extractor module.

[0031] In particular, the pre-training method is self-supervised and optimized using the global mutual information loss function, where the calculation formula of the global mutual information loss function is as follows: in, is the ReLU function used to approximate the gradient, is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; The objective function of the feature extractor module in the pre-training phase is as follows: pass Calculate the loss value and calculate and update the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module .

[0032] During formal training, the training of the regressor module and the training of the feature extractor module are carried out simultaneously. The loss function formula during the training process is as follows: in, Indicates the true value of temperature, represents the temperature detection value output by the regressor module, for The segmentation points of the function; In the late stage of converter steelmaking during formal training, the objective function of the furnace temperature detection model is as follows: The function combines the advantages of the MSEloss function and the MAEloss function. The function calculates the loss value and calculates and updates the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm. The network parameters of the regressor module , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module The network parameters of the regressor module .

[0033] The feature extractor module, the regressor module, and the output module are communicatively connected in sequence, the improved attention module is included in the feature extractor module, and the training and learning module is communicatively connected with the feature extractor module and the regressor module respectively.

[0034] Example 5 The feature extractor module automatically extracts features from the furnace mouth flame data to obtain feature data. During the feature extraction process of the feature extractor module, the improved ECA module in the feature extractor module performs 1D convolution operation on the channel dimension to perform weighted calculation on different channel weights to improve the validity of the feature data. The regressor performs regression calculation based on the feature data calculated by the feature extractor, outputs the normalized dimensionless temperature detection value, and then obtains the temperature detection value with temperature dimension through the output module. The training and learning module updates the parameters of the feature extractor module and the regressor module, and continuously iterates to improve the performance of the feature extractor module and the regressor module.

[0035] like Figure 2 As shown, the embodiment of the present invention provides a method for detecting the temperature in the late stage of converter steelmaking. The main process includes the construction of a data set, the establishment, training and testing of a model. Specifically, the method includes the following steps: Step 1: Collect several furnace mouth flame images in the late stage of converter steelmaking to construct a data set; Step 2: Preprocess the data set through normalization and data enhancement to obtain enhanced data; Step 3, define the temperature detection model in the late stage of converter steelmaking, including determining the input and output of the model, determining the feature extractor module, improved ECA module, regressor module, loss function and objective function in the model; Step 4, divide the data set into a training set, a validation set and a test set, use the training set to train the converter steelmaking late furnace temperature detection model, and use the validation set to verify the converter steelmaking late furnace temperature detection model; among them, 80% is the training set, 10% is the validation set, and 10% is the test set; Step 5: Input the test set into the furnace temperature detection model in the later stage of converter steelmaking, and output the prediction result of the furnace temperature detection.

[0036] Example 6 Based on Example 5, step 1 includes the following sub-steps: Step 1.1, record the furnace mouth flame video of the converter steelmaking in the later stage through an industrial camera; Step 1.2, select video clips with less image interference and good imaging quality to extract frames, and obtain several furnace mouth flame images; Step 1.3, crop the furnace flame image, remove the invalid area, and retain the valid area; Step 1.4: Correspond the cropped furnace mouth flame image to the furnace temperature to construct a data set.

[0037] Example 7 Based on Example 5 or 6, step 2 includes the following sub-steps: Step 2.1: Normalize the data. The furnace flame image is in RGB color space, and the color value range is 0 to 255. Map the color value range of the training set and the test set to the interval (0, 1). Step 2.2: Perform data enhancement, including randomly flipping and rotating the furnace flame image horizontally. The probability of random horizontal flipping is p=0.5, and the angle range of random rotation is set to (-15°, 15°).

[0038] Example 8 On the basis of Example 5, 6 or 7, the 1*1 convolution kernel size of the improved ECA module in step 3 is set to 3, that is, three similar channels are allowed to exchange information; the loss function of the feature extractor module in the pre-training stage is the maximum mutual information loss, and the loss function of the feature extractor module and the regressor module in the joint training stage is ; During the pre-training phase, the objective function of the feature extractor module is as follows: in, and Represents the intermediate feature data of the feature extractor module during the calculation process, and Indicates that after Layer and Layer Network Layer, It is an activation function used to approximate the gradient. Its shape is similar to the RELU function but smoother. is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; In the formal training phase, the objective function of the feature extractor module and the regressor module is trained together as follows: .

[0039] Example 9 Based on Embodiment 5, 6, 7 or 8, step 4 comprises the following sub-steps: Step 4.1, divide the data set into training set, validation set and test set, of which 80% is training set, 10% is validation set and 10% is test set; Step 4.2: Pre-train the feature extractor module, including 20 rounds of pre-warming and 300 rounds of formal training, where the batch size is 40, the adam optimizer is used, the initial learning rate for the 20 rounds of pre-warming is 0.0001, the initial learning rate for the 300 rounds of formal training is 0.001, the cosine annealing learning rate is used, and the loss function is the global mutual information loss; Step 4.3: Train the feature extractor module and the regressor module together, including 20 rounds of preheating and 300 rounds of formal training, where the batch size is 40, the adam optimizer is used, the initial learning rate for the 20 rounds of preheating is 0.0001, the initial learning rate for the 300 rounds of formal training is 0.002, the cosine annealing learning rate is used, and the loss function is Huberloss loss; Step 4.4: Use evaluation indicators to evaluate the temperature detection model in the late stage of converter steelmaking. The evaluation indicators include accuracy ( ), mean square error (MSE), root mean square error (RMSE) and mean absolute percentage error (MAPE), measure the average difference between the output value (detection value) of the converter steelmaking furnace temperature detection model and the true value. The difference between the model output value and the true value of the temperature is considered accurate if it is within 15°C; the greater the accuracy, the closer the output value of the converter steelmaking furnace temperature detection model is to the true value, and the smaller the other indicators, the closer the output value of the converter steelmaking furnace temperature detection model is to the true value, indicating that the prediction accuracy of the converter steelmaking furnace temperature detection model is higher; The formulas for accuracy, mean square error, root mean square error, and mean absolute percentage error are as follows: .

[0040] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A temperature detection model for the late stage of converter steelmaking, characterized in that: The model includes: The feature extractor module is a multi-layer structure based on a convolutional neural network, which is composed of a convolution and pooling layer, four block layers, and an average pooling layer connected in front and behind, and automatically extracts features from the furnace mouth flame image to obtain feature data; An improved ECA module, which is included in each block layer of the feature extractor module, performs a 1D convolution operation on the channel dimension to perform weighted calculation on different channel weights, thereby improving the validity of the feature data; A regressor module, the regressor module is connected to the feature extractor module, is a double-layer structure based on a fully connected neural network, calculates the feature data calculated by the feature extractor module to obtain a normalized dimensionless detection value; A training learning module, wherein the training learning module connects the feature extractor module and the regressor module; during pre-training, the training learning module trains the feature extractor module alone, and during formal training, the training learning module trains the feature extractor module and the regressor module at the same time, and performs parameter update through error back propagation; An output module is connected to the regressor module, and performs a denormalization operation on the dimensionless detection value output by the regressor module to obtain a temperature detection value with a temperature dimension. The smaller the error between the temperature detection value and the true value, the more accurate the temperature detection of the model is.

2. The temperature detection model in the late stage of converter steelmaking as claimed in claim 1, characterized in that: The parameters of the average pooling layer in the feature extractor module are adjusted according to the size characteristics of the input furnace flame image, and the formula is as follows: in, represents the height of the furnace flame image, Indicates the width of the furnace mouth flame image; For the pooling operation class, is an instantiated pooling operation class object. is the feature data before the pooling operation, Realize The pooling operation, is the feature data after the pooling operation, that is, the output of the feature extractor module.

3. The temperature detection model in the late stage of converter steelmaking as claimed in claim 2, characterized in that: The improved ECA module performs weighted calculation on different channels, and the formula is as follows: in, The characteristic data before the ECA module is improved. The feature data after passing the improved ECA module; The internal calculation process of the improved ECA module is as follows: in, For the pooling operation class, The function compresses the last dimension of the feature data. The function swaps the last two dimensions of the feature data; The function performs a 1D convolution operation on the last dimension of the feature data to achieve cross-channel information exchange. The function adds a dimension after the last dimension of the feature data. The function is the activation function, The function expands the feature data to the specified dimension; During the entire calculation process, The feature dimension changes are as follows: [B,C,H,W]->[B,C,1,1]->[B,C,1]->[B,1,C]->[B,1,C]->[B,C,1]->[B,C,1,1]->[B,C,H,W] Finally, the calculated attention Weight and multiply, thereby enhancing important features and suppressing unimportant features.

4. The temperature detection model in the converter steelmaking late stage as claimed in claim 3, characterized in that: During pre-training, the pre-training method is self-supervision and the global mutual information loss function is used for optimization, wherein the calculation formula of the global mutual information loss function is as follows: in, is the ReLU function used to approximate the gradient, is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; The objective function of the feature extractor module in the pre-training phase is as follows: pass Calculate the loss value, calculate and update the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module. .

5. The temperature detection model in the late stage of converter steelmaking as claimed in claim 4, characterized in that: During formal training, the training of the regressor module and the training of the feature extractor module are carried out simultaneously, and the loss function formula during the training process is as follows: in, Indicates the true value of temperature, represents the temperature detection value output by the regressor module, for The segmentation points of the function; In the late stage of converter steelmaking during formal training, the objective function of the furnace temperature detection model is as follows: pass The function calculates the loss value and calculates and updates the network parameters of the feature extractor module through back propagation and stochastic gradient descent optimization algorithm. The network parameters of the regressor module , the update formula is as follows: in, is the learning rate, is the gradient operator; This process is continuously performed to optimize the network parameters of the feature extractor module. The network parameters of the regressor module .

6. A method for detecting the temperature in a converter steelmaking furnace in the later stage, characterized in that: The method comprises the following steps: Step 1: Collect several furnace mouth flame images in the late stage of converter steelmaking to construct a data set; Step 2: preprocessing the data set by normalization and data enhancement to obtain enhanced data; Step 3, define the temperature detection model in the late stage of converter steelmaking, including determining the input and output of the model, determining the feature extractor module, improved ECA module, regressor module, loss function and objective function in the model; Step 4, dividing the data set into a training set, a validation set and a test set, using the training set to train the converter steelmaking late furnace temperature detection model, and using the validation set to verify the converter steelmaking late furnace temperature detection model; wherein 80% is the training set, 10% is the validation set, and 10% is the test set; Step 5: input the test set into the converter steelmaking late furnace temperature detection model, and output the prediction result of the furnace temperature detection.

7. The method for detecting the temperature in the converter steelmaking late stage as claimed in claim 6, characterized in that: The step 1 includes the following sub-steps: Step 1.1, record the furnace mouth flame video of the converter steelmaking in the later stage through an industrial camera; Step 1.2, select video clips with less image interference and good imaging quality to extract frames, and obtain several furnace mouth flame images; Step 1.3, cropping the furnace mouth flame image, removing invalid areas, and retaining valid areas; Step 1.4, the cropped furnace mouth flame image is matched with the temperature inside the furnace to construct the data set.

8. The method for detecting the temperature in the converter steelmaking late stage as claimed in claim 6, characterized in that: The step 2 includes the following sub-steps: Step 2.1, perform data normalization, the furnace mouth flame image is RGB color space, the color value range is 0 to 255, and the color value range of the training set and the test set is mapped to the (0, 1) interval; Step 2.2, perform data enhancement, including randomly flipping and rotating the furnace mouth flame image horizontally, wherein the probability of random horizontal flipping p=0.5, and the angle range of random rotation is set to (-15°, 15°).

9. The method for detecting the temperature in the converter steelmaking late stage as claimed in claim 6, characterized in that: The 1*1 convolution kernel size of the improved ECA module in step 3 is set to 3, that is, three similar channels are allowed to exchange information; the loss function of the feature extractor module in the pre-training stage is the maximum mutual information loss, and the loss function of the feature extractor module and the regressor module in the joint training stage is ; In the pre-training stage, the objective function of the feature extractor module is as follows: in, and represents the intermediate feature data of the feature extractor module during the calculation process, and Indicates that after Layer and Layer Network Layer, It is an activation function used to approximate the gradient. Its shape is similar to the RELU function but smoother. is a discriminator constructed by a neural network, used to judge Is it Calculated by neural network, for and The approximate value of the negative entropy of the joint distribution of for and The approximation of the negative entropy of the marginal distribution of is the global mutual information loss; In the formal training phase, the objective function of the feature extractor module and the regressor module for joint training is as follows: 。 10. The method for detecting the temperature in the converter steelmaking late stage according to claim 6, characterized in that: The step 4 includes the following sub-steps: Step 4.1, divide the data set into a training set, a validation set and a test set, where 80% is the training set, 10% is the validation set, and 10% is the test set; Step 4.2, pre-training the feature extractor module, including 20 rounds of preheating and 300 rounds of formal training, wherein the batch size is 40, the adam optimizer is used, the initial learning rate of the 20 rounds of preheating is 0.0001, the initial learning rate of the 300 rounds of formal training is 0.001, the cosine annealing learning rate is used, and the loss function is the global mutual information loss; Step 4.3, the feature extractor module and the regressor module are jointly trained, including 20 rounds of preheating and 300 rounds of formal training, wherein the batch size is 40, the adam optimizer is used, the initial learning rate of the 20 rounds of preheating is 0.0001, the initial learning rate of the 300 rounds of formal training is 0.002, the cosine annealing learning rate is used, and the loss function is Huberloss loss; Step 4.4, use evaluation indicators to evaluate the converter steelmaking later furnace temperature detection model, the evaluation indicators include accuracy, mean square error, root mean square error and mean absolute percentage error, measure the average difference between the output value of the converter steelmaking later furnace temperature detection model and the true value, the difference between the model output value and the true value of the temperature is within 15 ° C. The formulas for accuracy, mean square error, root mean square error and mean absolute percentage error are as follows: 。

Images