Iron-like image-based blast furnace molten iron silicon content estimation method and system

Abstract

The present application provides a kind of based on iron sample image blast furnace molten iron silicon content estimation method and system, it is related to blast furnace ironmaking technical field, the method includes collecting blast furnace field iron sample image, calibration according to silicon content;Iron sample image data set is constructed, and is divided into training set, verification set and test set;Data set is handled using data enhancement strategy;The training set and verification set after enhancement are used to train blast furnace molten iron silicon content estimation regression model;Evaluation index is designed, and the performance of model is verified using the test set after enhancement;Field iron sample image is input into trained model, and silicon content is predicted.The present application realizes the silicon content prediction of high precision, real-time by applying computer vision technology to blast furnace molten iron silicon content estimation, while reducing cost, enhancing environmental adaptability, and solving the problem of limited image quantity and uneven label distribution.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of blast furnace ironmaking, and particularly relates to a method and system for estimating silicon content of molten iron in a blast furnace based on an iron sample image. BACKGROUND

[0002] As a core link of the steel industry, the production efficiency and product quality of blast furnace ironmaking directly relate to the competitiveness of the entire industry. The thermal state of the blast furnace is a key parameter in the smelting process, directly determining the quality of molten iron, fuel consumption, and overall production efficiency. However, the extreme working environment inside the blast furnace, including high temperature, high pressure, and strong corrosion, makes it extremely difficult to directly measure the thermal state inside the blast furnace.

[0003] In order to indirectly reflect the thermal state inside the blast furnace, the industry generally uses the silicon content of molten iron as an indicator. The silicon content of molten iron is closely related to the furnace temperature, and too high or too low silicon content will adversely affect the quality and flowability of molten iron, and thus affect the subsequent smelting and processing process. Therefore, accurately controlling the silicon content of molten iron becomes a core task for adjusting the thermal state of the blast furnace.

[0004] Currently, the methods for determining silicon content mainly include offline measurement, online measurement, mechanism model prediction, and data-driven model prediction.

[0005] Offline measurement: This method measures the silicon content by sampling and testing the molten iron. Although it has high accuracy, it takes a long time, usually 1 to 1.5 hours, leading to time lag and difficulty in timely adjusting the blast furnace operating conditions.

[0006] Online measurement: Based on the laser-induced breakdown spectroscopy (LIBS) technology, this method uses a spectrometer installed on site to monitor the silicon content of the molten iron surface in real time. However, this method has high equipment cost and is easily affected by harsh working conditions such as molten iron level fluctuations, dust, and high temperature radiation. In addition, the processing of spectral signals and quantitative analysis models still need to be further improved.

[0007] Mechanism model prediction: Based on the chemical reactions in blast furnace smelting, this method derives a mechanism model of silicon content based on principles such as heat balance and mass conservation. However, the blast furnace ironmaking process is complex and variable, involving numerous random phenomena and disturbances, making it difficult for mechanism models to accurately describe the actual process, and the simulation results are often distorted.

[0008] Data-driven model prediction method: a prediction model is established using historical process data captured by sensors, mainly using statistical methods and machine learning techniques. Although this method has great potential, it is highly dependent on the quality and quantity of data, and the continuity, dynamics, nonlinearity and multi-scale effects of the ironmaking process pose great challenges to model selection and construction. In addition, the accuracy of some process variables is easily affected by the environment when the furnace condition is unstable, further increasing the difficulty of model learning.

[0009] In view of the limitations of existing methods in terms of measurement accuracy, real-time performance, cost and environmental adaptability, there is an urgent need to develop a more accurate, efficient, economical and adaptable method for estimating the silicon content of molten iron to better guide the blast furnace ironmaking process and improve production efficiency and product quality. SUMMARY

[0010] To this end, the embodiments of the present application provide a blast furnace molten iron silicon content estimation method and system based on iron sample images, which are used to solve the problems of low precision, poor real-time performance, high cost and weak environmental adaptability in measuring the silicon content of molten iron in the prior art.

[0011] To solve the above problems, the embodiments of the present application provide a blast furnace molten iron silicon content estimation method based on iron sample images, which comprises:

[0012] Step S1: Collecting iron sample images of the blast furnace site, and labeling the corresponding iron sample images according to the silicon content of each iron sample;

[0013] Step S2: Based on the labeled iron sample images, an iron sample image dataset is constructed, and the iron sample image dataset is divided into a training set, a validation set and a test set according to a set proportion;

[0014] Step S3: Using a set data enhancement strategy, the iron sample image dataset is subjected to data enhancement processing;

[0015] Step S4: Using the image-enhanced training set and validation set, a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images is trained to obtain a trained blast furnace molten iron silicon content estimation regression model based on iron sample images;

[0016] Step S5: Designing evaluation indicators, the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images is verified using the image-enhanced test set;

[0017] Step S6: Inputting the iron sample images collected on site into the trained blast furnace molten iron silicon content estimation regression model based on iron sample images to obtain the silicon content of the iron sample corresponding to the iron sample image.

[0018] Preferably, the data enhancement strategy includes an automatic contrast color adjustment transformation strategy, a geometric transformation, and a Cutout transformation.

[0019] Preferably, the method for training the pre-constructed image-based blast furnace molten iron silicon content estimation regression model using the image-enhanced training set and validation set includes:

[0020] First, the parameters of the pre-constructed image-based blast furnace molten iron silicon content estimation regression model are initialized; then, the image-enhanced training set is input into the model for forward propagation calculation and to obtain the prediction result; next, the constructed weighted penalty loss function is used to calculate the error between the model prediction result and the true silicon content, and the network parameters are updated through back propagation; during the training process, the image-enhanced validation set is used regularly to evaluate the performance of the model, and the hyperparameters of the model are adjusted according to the evaluation result; after multiple iterations of training, the network parameters are continuously optimized until the preset number of iterations is reached, at which point the model training is completed, and the trained image-based blast furnace molten iron silicon content estimation regression model is finally obtained.

[0021] Preferably, the network structure of the pre-constructed image-based blast furnace molten iron silicon content estimation regression model is:

[0022] The first layer is a scaling layer, which first scales the image to (1, 224, 224) to meet the structured operation of the feature extraction network;

[0023] The second layer is a convolution and pooling layer, which first performs a convolution operation with a kernel size of 7x7, a stride of 2, and a padding of 3, and outputs a feature map with a size of (64, 112, 112), where 64 represents the number of channels of the feature map and 112x112 represents the size of each feature map; secondly, a maximum pooling operation is performed with a pooling window size of 3x3, a stride of 2, and a padding of 1, and outputs a feature map with a size of (64, 56, 56), where 64 represents the number of channels of the feature map and 56x56 represents the size of each feature map;

[0024] The third layer to the tenth layer are residual layers, and the residual layers are composed of convolution, a multi-scale attention module and a residual connection. In one branch, a 3*3 convolution operation is first performed with a step length of 1, wherein the step length of the fifth layer, the seventh layer and the ninth layer is 2 and the padding is 1, then batch normalization and a ReLU activation function are performed, and then a 3*3 convolution with a step length of 1 and padding of 1 is performed, and the batch normalization is performed to input the multi-scale attention module; the multi-scale attention module divides the input features into three branches, wherein two branches are spliced after global pooling in the X and Y directions, and then the channel fusion is performed through a 1*1 convolution to obtain global features; the third branch captures local features through a 3*3 convolution layer; after obtaining two different scale feature representations, corresponding weights are generated, and matrix multiplication calculation is performed with the features of another scale, that is, the weights generated by the features of different scales are used to adjust the features of the same scale, and finally the global spatial attention representation under the current scale is obtained; finally, the spatial attentions of two different scales are added, and the weights are generated through a sigmoid function to adjust the input and obtain the final output; in the second branch of the residual layer, the input features are directly kept unchanged; after the features of the two branches are added, the output features are finally output through a ReLU activation function;

[0025] The eleventh layer performs a global average pooling operation, and the output feature map size is (512, 1, 1);

[0026] The twelfth layer performs feature calibration work, keeps the feature map size unchanged, and finally obtains a feature vector with a size of 512 for representing the iron sample silicon content;

[0027] The thirteenth layer uses a fully connected layer to map the (512, 1, 1) feature map to a (1, 1, 1) silicon content value, and finally realizes the estimation task of silicon content.

[0028] Preferably, a feature calibration module is introduced in the twelfth layer, which is used to smooth the feature distribution between similar silicon content samples, so as to adjust the extracted feature map.

[0029] Preferably, the feature calibration module is used to smooth the feature distribution between similar silicon content samples, and specifically includes:

[0030] After receiving the features of the previous layer, first, the samples are sorted according to the silicon content label from small to large, and are uniformly divided into multiple intervals; wherein for each silicon content interval, as shown in table b, arranged in ascending order of silicon content, the feature mean μ of the interval is calculated by the following formula b and the standard deviation σ b :

[0031]

[0032]

[0033] where N b denotes the number of samples in the b-th silicon content interval, z i is the feature vector of the i-th sample in the b-th silicon content interval, and denotes the element-wise multiplication operation.

[0034] Then, the feature information of adjacent intervals is smoothed by a symmetric kernel function, so as to obtain a smoother feature distribution. The smoothed feature mean and standard deviation are calculated by the following formula

[0035]

[0036]

[0037] where k(b, b') is a symmetric kernel function measuring the similarity between the b-th and b'-th intervals, and B denotes the index set of all intervals in the silicon content range; μ b′ denotes the feature mean of the b'-th silicon content interval, and σ b′ denotes the feature standard deviation of the b'-th silicon content interval.

[0038] Finally, the feature is calibrated by standardization and recoloring.

[0039] z i ' = (z i - μ b ) / σ b ;

[0040]

[0041] where z i ' denotes the standardized feature vector of the i-th sample in the b-th silicon content interval, denotes the calibrated feature vector of the i-th sample in the b-th silicon content interval, i = 1,..., N b , and b ∈ B.

[0042] Preferably, the weighted penalty loss function is:

[0043]

[0044] where Loss denotes the weighted penalty loss function, N denotes the total number of samples, y i denotes the true silicon content value of the i-th sample, denotes the estimated silicon content value of the sample, and α and β are penalty factors.

[0045] ​The evaluation indexes include root mean square error (RMSE), mean absolute error (MAE) and cumulative score (CS);

[0046] The calculation formula of the root mean square error (RMSE) is:

[0047]

[0048] The calculation formula of the mean absolute error (MAE) is:

[0049]

[0050] The calculation formula of the cumulative score (CS) is:

[0051]

[0052]

[0053] In the formula, H represents the actual silicon content of the sample t, represents the predicted silicon content output by the model, N Si represents the total number of samples, H 0.1 (t) and H 0.05 (t) is a Heaviside step function.

[0054] Preferably, the Heaviside step function is defined as follows:

[0055]

[0056] In the formula, H x (t) is a Heaviside step function, and x is an independent variable.

[0057] The embodiment of the present application also provides a blast furnace molten iron silicon content estimation system based on an iron sample image, which is used to realize the blast furnace molten iron silicon content estimation method based on an iron sample image as described above, and specifically comprises:

[0058] An iron sample image collection and calibration module is used to collect iron sample images of a blast furnace site, and calibrate the corresponding iron sample images of each iron sample according to the silicon content of the iron sample;

[0059] A data set division module is used to construct an iron sample image data set based on the calibrated iron sample images, and divide the iron sample image data set into a training set, a verification set and a test set according to a set proportion;

[0060] A data enhancement module is used to perform data enhancement processing on the iron sample image data set by using a set data enhancement strategy;

[0061] a model training module, configured to train a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images by using the image-enhanced training set and the image-enhanced verification set, and obtain a trained blast furnace molten iron silicon content estimation regression model based on iron sample images;

[0062] a model performance verification module, configured to design an evaluation index and verify the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images by using an image-enhanced test set;

[0063] a silicon content prediction module, configured to input an iron sample image collected on site into the trained blast furnace molten iron silicon content estimation regression model based on iron sample images, and obtain the silicon content of the iron sample corresponding to the iron sample image.

[0064] As can be seen from the above technical solutions, the present application has the following beneficial effects:

[0065] (1) The present application first introduces computer vision technology into the field of silicon content estimation in the blast furnace ironmaking process. By converting this practical problem into a deep regression task, the present application can automatically extract features from a large number of iron sample images and learn the target value of the silicon content, thereby achieving accurate prediction of the silicon content.

[0066] (2) Compared with traditional methods, the present application completely relies on iron sample images for silicon content estimation without the need for human intervention or the support of complex sensor equipment. This innovation not only reduces the dependence on external resources, but also significantly improves the estimation speed, providing more timely and effective operation guidance for on-site operators.

[0067] (3) In view of the relatively limited number of iron sample images, the present application designs an efficient data enhancement method to generate diverse training data, thereby enhancing the robustness and prediction accuracy of the model. In addition, in view of the actual situation of label distribution imbalance in the data set, the present application optimizes the network structure and loss function, effectively alleviating the negative effects of label distribution imbalance, and further improving the overall performance of the model in the silicon content estimation task. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments or the prior art, the drawings needed in the embodiments will be briefly described below. The features and advantages of the present application can be more clearly understood by referring to the drawings. The drawings are schematic and should not be construed as any limitation on the present application. Those skilled in the art can obtain other drawings without creative labor on the basis of these drawings. Among them:

[0069] Figure 1 A flowchart of a blast furnace molten iron silicon content estimation method based on iron sample images provided in the embodiments;

[0070] Figure 2 is a schematic diagram of on-site sampling in the embodiment;

[0071] Figure 3 is a structural schematic diagram of a feature calibration module in the embodiment;

[0072] Figure 4 is a scatter plot of estimated silicon content and actual silicon content in the embodiment;

[0073] Figure 5 is a block diagram of a blast furnace molten iron silicon content estimation system based on iron sample images provided in the embodiment. DETAILED DESCRIPTION

[0074] To make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0075] Embodiment one

[0076] To solve the problems of low precision, poor real-time performance, high cost and weak environmental adaptability in measuring the silicon content of molten iron in the prior art, as shown in the formula (1), the embodiments of the present application propose a blast furnace molten iron silicon content estimation method based on iron sample images, which comprises the following steps: Figure 1

[0077] Step S1: Collecting iron sample images on the spot of a blast furnace, and labeling the corresponding iron sample images according to the silicon content of each iron sample;

[0078] Step S2: Based on the labeled iron sample images, constructing an iron sample image dataset, and dividing the iron sample image dataset into a training set, a validation set and a test set according to a set proportion;

[0079] Step S3: Using a set data enhancement strategy to perform data enhancement processing on the iron sample image dataset;

[0080] Step S4: Using the image-enhanced training set and validation set to train a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images, to obtain a trained blast furnace molten iron silicon content estimation regression model based on iron sample images;

[0081] Step S5: Designing an evaluation index, and using the image-enhanced test set to verify the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images; ​

[0082] Step S6: Input the on-site collected iron sample image into the trained regression model for estimating silicon content in blast furnace hot metal based on the iron sample image to obtain the silicon content of the iron sample corresponding to the iron sample image.

[0083] As can be seen from the above technical solution, this invention proposes a method for estimating the silicon content of blast furnace molten iron based on iron sample images. By collecting and calibrating on-site iron sample images of the blast furnace, a dataset is constructed, and data augmentation strategies are applied. A deep regression model is trained using training and validation sets to automatically extract image features and learn the target value of silicon content. This method is the first to apply computer vision technology to the estimation of silicon content in blast furnace ironmaking, overcoming the problems of traditional methods such as reliance on manual labor or complex equipment, low accuracy, and poor real-time performance. It has significant advantages such as high speed, low cost, and strong environmental adaptability. Furthermore, addressing the challenges of a limited number of images and imbalanced label distribution, this invention designs a specialized data augmentation method and an optimized network structure, improving the robustness and accuracy of the model.

[0084] In step S1, during the blast furnace tapping operation, the on-site operator will use a sampling spoon to collect molten iron samples from the molten iron trough, such as... Figure 2 As shown, when molten iron comes into contact with the cooler sampling spoon, it quickly solidifies into a thin sheet, forming an iron sample. To obtain images of the iron samples, a specialized imaging box is used, and images are taken with a camera using fixed parameters under constant color temperature and brightness conditions. This imaging method minimizes the influence of external factors on the color of the iron sample images, ensuring image quality. Subsequently, these iron samples are sent to a laboratory for analysis to determine their accurate silicon content. Based on the silicon content results of each iron sample, its corresponding image is calibrated. Throughout the entire process, all iron sample images are taken at the blast furnace site using a controlled light box to ensure consistent lighting conditions for all images.

[0085] In step S2, based on the calibrated iron sample images, an iron sample image dataset is constructed, and the iron sample image dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0086] In step S3, to ensure the robustness of the trained model, a large number of iron sample images containing known silicon content are needed as a dataset. However, in reality, only one iron sample can be provided per furnace of molten iron, and these samples need to be photographed on-site, which undoubtedly increases the difficulty of obtaining these images. To overcome this challenge, this embodiment designs a novel data augmentation technique aimed at generating more diverse training data.

[0087] Further, the embodiment introduces an automatic contrast color adjustment strategy. This strategy can simulate various lighting conditions that may be encountered during live shooting, by enhancing or weakening the light intensity in the image, to help the model better identify features under different lighting conditions. In addition to color adjustment, the embodiment also uses geometric transformation techniques, including ShearX, ShearY, TranslateX, and TranslateY transformation methods, to cope with changes in camera angle and possible slight misalignment during shooting. These geometric transformations can simulate natural changes, allowing the model to identify image patterns in different spatial directions. In addition, the embodiment introduces random occlusion areas in the image through Cutout transformation to improve the model's detection ability under incomplete feature conditions.

[0088] In step S4, since the color depth of the iron sample as a whole and the local golden yellow spots on it have important visual indications for predicting the silicon content, it is particularly important to extract multi-scale features in the process of improving the prediction accuracy of the model. The model network structure proposed in this embodiment is designed based on this requirement. The network structure of the blast furnace molten iron silicon content estimation regression model based on iron sample images is as follows:

[0089] The first layer is a scaling layer, which first scales the image to (1, 224, 224) to meet the structured operation of the feature extraction network.

[0090] The second layer is a convolution and pooling layer. First, a convolution operation with a kernel size of 7x7, a stride of 2, and a padding of 3 is performed, and the output feature map size is (64, 112, 112), where 64 represents the number of channels of the feature map and 112x112 is the size of each feature map. Then, a maximum pooling operation with a window size of 3x3, a stride of 2, and a padding of 1 is performed, and the output feature map size is (64, 56, 56), where 64 represents the number of channels of the feature map and 56x56 is the size of each feature map.

[0091] Layers three through ten are residual layers, each consisting of convolutions, a multi-scale attention module, and residual connections. Within each branch, a 3×3 convolution with a stride of 1 is performed first, with layers five, seven, and nine having a stride of 2 and padding of 1. This is followed by batch normalization and ReLU activation. Afterward, the input is processed by another 3×3 kernel with a stride of 1 and padding of 1, and then batch normalized before being fed into the multi-scale attention module. The multi-scale attention module divides the input features into three branches. Two of these branches undergo global pooling in the X and Y directions, are concatenated, and then fused using a 1×1 convolution to obtain global features. The third branch uses a 3×3 convolutional layer to capture local features. After obtaining feature representations at two different scales, corresponding weights are generated for each scale, and matrix multiplication is performed with the features at the other scale. This uses weights generated from features at different scales to adjust the current feature representation, ultimately obtaining the global spatial attention representation at the current scale. Finally, the two spatial attention representations at different scales are added, and weights are generated using the sigmoid function to adjust the input and obtain the final output. The feature dimensions remain unchanged after passing through the attention module. In the second branch of the residual layer, the input features are kept unchanged; the features from the two branches are added together, and finally, a ReLU activation function is applied to output the features.

[0092] The eleventh layer performs a global average pooling operation, and the output feature map size is (512, 1, 1).

[0093] The twelfth layer performs feature calibration, keeping the feature map size unchanged, and finally obtains a feature vector of size 512 to represent the silicon content of the iron sample.

[0094] The thirteenth layer uses a fully connected layer to map the feature map of (512,1,1) to the silicon content value of (1,1,1), ultimately achieving the task of estimating the silicon content.

[0095] The relatively small number of samples with high and low silicon content leads to data imbalance, making feature statistics more susceptible to noise interference and resulting in significant differences in feature statistics between samples with adjacent silicon content. To address this challenge, this embodiment innovatively introduces a feature calibration module at the twelfth layer of the network structure. The main function of this module is to smooth the feature distribution among samples with similar silicon content, thereby enabling fine-tuning of the extracted feature maps.

[0096] Furthermore, such as Figure 3 As shown, after receiving the features from the previous layer, the samples are first sorted according to the silicon content label from smallest to largest, and then evenly divided into multiple intervals. For each silicon content interval, as shown in Table b, the features mean μ of the interval is calculated using the following formula, arranged in ascending order of silicon content.b and standard deviation σ b :

[0097]

[0098]

[0099] where N b denotes the number of samples in the b-th silicon content interval, z i is the feature vector of the i-th sample in the b-th silicon content interval, and denotes the element-wise multiplication operation.

[0100] Then, the feature information of adjacent intervals is smoothed by a symmetric kernel function, so as to obtain a smoother feature distribution. The smoothed feature mean and standard deviation are calculated by the following formula and standard deviation

[0101]

[0102]

[0103] where k(b, b') is a symmetric kernel function measuring the similarity between the b-th and b'-th intervals, B denotes the index set of all intervals in the silicon content range; μ b′ denotes the feature mean of the b'-th silicon content interval, σ b′ denotes the feature standard deviation of the b'-th silicon content interval.

[0104] Finally, the feature is calibrated by standardization and recoloring;

[0105] z i ' = (z i - μ b ) / σ b ;

[0106]

[0107] where z i ' denotes the standardized feature vector of the i-th sample in the b-th silicon content interval, denotes the calibrated feature vector of the i-th sample in the b-th silicon content interval, i = 1,..., N b , b ∈ B.

[0108] With the above method, the embodiment successfully realizes smoothing processing in the feature space of the iron sample image, so that samples with similar silicon content are closer in feature representation. This improvement not only reduces the interference of inherent noise by passing the features of adjacent intervals, but also enables the model to learn more reliable and consistent feature representations, especially when dealing with insufficient representative iron sample features, significantly improving the model's performance on unbalanced data sets.

[0109] Next, the loss function design of the embodiment is introduced in detail. For the specific goal of silicon content estimation task, the loss function designed by the embodiment is composed of multiple components. First, the mean square error (MSE) term is introduced to quantify the squared difference between the true silicon content and the model estimated value. Since it is difficult to achieve absolute accurate estimation in practical application, the hit rate within 0.1% or 0.05% of absolute error is usually used as the evaluation standard. Therefore, the embodiment adds two additional penalty terms in the loss function, corresponding to the error thresholds of 0.1% and 0.05% respectively. Among them, the 0.1% threshold is set as the maximum acceptable error of silicon content estimation, while the 0.05% threshold aims to encourage the model to pursue higher prediction accuracy. For the estimated values exceeding these thresholds, the corresponding penalties will be applied to guide the model to continuously improve the accuracy of prediction.

[0110] Combined with the effective label density and the above penalty threshold, the embodiment constructs a complete weighted penalty loss function, which is specifically as follows:

[0111]

[0112] In the formula, Loss represents the weighted penalty loss function, N represents the total number of samples, y i represents the true silicon content value of the i-th sample, represents the estimated silicon content value of the sample, and α and β are penalty factors.

[0113] Further, a method for training a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images using the enhanced image training set and validation set includes:

[0114] First, the parameters of the pre-constructed regression model for estimating the silicon content of molten iron in a blast furnace based on iron sample images are initialized; then, the training set after image enhancement is input into the model for forward propagation calculation, and the prediction result is obtained; next, the weighted penalty loss function is used to calculate the error between the model prediction result and the true silicon content, and the network parameters are updated through back propagation; during the training process, the performance of the model is evaluated regularly using the image-enhanced validation set, and the hyperparameters of the model are adjusted according to the evaluation results; after multiple iterations of training, the network parameters are continuously optimized until the preset number of iterations is reached, at which point the model training is complete, and the trained regression model for estimating the silicon content of molten iron in a blast furnace based on iron sample images is obtained.

[0115] In step S5, in order to comprehensively and accurately evaluate the performance of the model proposed in this embodiment, this embodiment adopts root mean square error (RMSE), mean absolute error (MAE), and cumulative score (CS) as core evaluation indicators. In actual production applications, when the absolute error between the model's predicted silicon content and the actual silicon content is within the range of 0.1% or 0.05%, it is generally considered that such prediction is accurate and reliable. Therefore, this embodiment particularly introduces the CS indicator to intuitively reflect the model's prediction ability within this critical error range. The specific calculation formulas of RMSE, MAE, and CS are as follows:

[0116]

[0117]

[0118]

[0119]

[0120] wherein, represents the actual silicon content of sample t, represents the predicted silicon content output by the model, N Si represents the total number of samples, H 0.1 (t) and H 0.05 (t) is the Heaviside step function, defined as follows:

[0121]

[0122] wherein, H x (t) is the Heaviside step function, and x is the independent variable.

[0123] In step S6, the iron sample images collected on site are input into the trained regression model for estimating the silicon content of molten iron in a blast furnace based on iron sample images, and the silicon content of the iron sample corresponding to the iron sample image is obtained.

[0124] To further illustrate the advantages of the present application, specific experiments are described below.

[0125] 1. Data Preparation and Experimental Platform

[0126] At the blast furnace operation site, we took strict measures to collect iron samples and take photos. All images were taken on site by controlling the light box, which was done to ensure that the lighting conditions of all photos remained highly consistent. At the same time, a camera with fixed parameters was used for shooting to minimize external factors that could affect the color of the iron sample images. A total of 2040 iron sample images were successfully collected, covering a wide range of silicon content from 0.06% to 0.77%.

[0127] To effectively evaluate and optimize our silicon content estimation regression model, we divided the dataset into training, validation, and test sets in the ratio of 8:1:1. The model was built based on the PyTorch deep learning framework, with an initial learning rate of 10^-3 (i.e. 0.001) and the AdamW optimizer was selected to dynamically adjust the learning rate to achieve better training results.

[0128] During training, we used a batch size of 32 and trained for 100 rounds. All these computing tasks were completed on a high-performance computer equipped with a 13th generation Intel Core i7-13700K processor, 32GB of RAM, and an RTX 3070 GPU with 8GB of video memory. Such hardware configuration provides strong support for our model training.

[0129] 2. Validate Model Performance

[0130] After the model training is completed, the performance of the model is verified on the test set. RMSE, MAE, CS-0.1, and CS-0.05 are used as evaluation criteria. The experimental results show that the model's RMSE reaches 0.0583, MAE reaches 0.0441, CS-0.05 reaches 66.02%, and CS-0.1 reaches 92.71%. The silicon content error range within 0.1% is effective for blast furnace operation, so it can be seen that our model has achieved a good result.

[0131] To visually display the relationship between estimated silicon content and true silicon content, we plotted a scatter plot, where the x-axis represents the actual silicon content and the y-axis represents the estimated silicon content. As shown in Figure 2, the estimated silicon content is very close to the actual silicon content, indicating that our model has achieved a good result. Figure 4As shown, the scatter points are concentrated near the y=x line, and most of the errors are within the range of [-0.1, 0.1]. Even in extreme cases where the silicon content is high or low (which is difficult to estimate), the model still maintains high accuracy. This consistency across different silicon content levels highlights the robustness and reliability of our method in real-time silicon content estimation.

[0132] Embodiment Two

[0133] As Figure 5 shown, the present application provides a blast furnace molten iron silicon content estimation system based on iron sample images, which is used to implement the blast furnace molten iron silicon content estimation method based on iron sample images in Embodiment One, and specifically includes:

[0134] An iron sample image collection and calibration module 100 is used to collect iron sample images on the blast furnace site and calibrate the iron sample images according to the silicon content of each iron sample;

[0135] A data set division module 200 is used to construct an iron sample image data set based on the calibrated iron sample images, and divide the iron sample image data set into a training set, a validation set, and a test set according to a set proportion;

[0136] A data enhancement module 300 is used to perform data enhancement processing on the iron sample image data set using a set data enhancement strategy;

[0137] A model training module 400 is used to train a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images using the image-enhanced training set and validation set, to obtain a trained blast furnace molten iron silicon content estimation regression model based on iron sample images;

[0138] A model performance verification module 500 is used to design evaluation indicators and verify the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images using the image-enhanced test set;

[0139] A silicon content prediction module 600 is used to input the iron sample images collected on site into the trained blast furnace molten iron silicon content estimation regression model based on iron sample images, to obtain the silicon content of the iron sample corresponding to the iron sample image.

[0140] The silicon content estimation system of molten iron based on the iron sample image of the embodiment is used for realizing the silicon content estimation method of molten iron based on the iron sample image, and therefore the specific embodiments in the silicon content estimation system of molten iron based on the iron sample image can be seen from the foregoing embodiment part of the silicon content estimation method of molten iron based on the iron sample image, for example, the iron sample image collection and calibration module 100, the data set division module 200, the data enhancement module 300, the model training module 400, the model performance verification module 500, and the silicon content prediction module 600 are respectively used for realizing steps S1, S2, S3, S4, S5, and S6 in the foregoing silicon content estimation method of molten iron based on the iron sample image, and therefore the specific embodiments can be referred to the description of the respective embodiment part, and details are not described herein again to avoid redundancy.

[0141] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt a computer program product in the form of being implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program codes.

[0142] The present application is described with reference to flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a means for implementing the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1 The means for implementing the functions specified in one flow or multiple flows and / or blocks.

[0143] These computer program instructions can also be stored in a computer readable storage medium capable of guiding the computer or other programmable data processing devices to work in a specific manner, so that the instructions stored in the computer readable storage medium produce a product including instruction means, which implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one flow or multiple flows and / or blocks Figure 1The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the operations specified in the flowchart block or blocks. Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the operations specified in the flowchart block or blocks. Figure 1 The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the operations specified in the flowchart block or blocks.

[0144] Obviously, the above-described embodiments are only examples for clearly illustrating the present application and are not intended to limit the present application. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, it is not necessary and also impossible to enumerate all the embodiments. The obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A method for estimating the silicon content of molten iron in a blast furnace based on an iron-like image, characterized by, The method comprises the following steps: Step S1: Collecting iron sample images on the spot of a blast furnace, and labeling the corresponding iron sample images according to the silicon content of each iron sample; Step S2: Based on the labeled iron sample images, constructing an iron sample image data set, and dividing the iron sample image data set into a training set, a validation set and a test set according to a set proportion; Step S3: Using a set data enhancement strategy to perform data enhancement processing on the iron sample image data set; Step S4: Using the training set and the validation set after image enhancement to train a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images, and obtaining a trained blast furnace molten iron silicon content estimation regression model based on iron sample images; Step S5: Designing an evaluation index, and verifying the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images by using the test set after image enhancement; Step S6: Inputting the iron sample images collected on the spot into the trained blast furnace molten iron silicon content estimation regression model based on iron sample images, and obtaining the silicon content of the iron sample corresponding to the iron sample image.

2. The iron-like image-based blast furnace molten iron Si content estimation method according to claim 1, characterized by, The data enhancement strategy comprises an automatic contrast color adjustment transformation strategy, geometric transformation and Cutout transformation.

3. The iron-like image-based blast furnace molten iron Si content estimation method according to claim 1, characterized by, The method for training the pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images by using the training set and the validation set after image enhancement comprises the following steps: Firstly, the parameters of the pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images are initialized; then, the training set after image enhancement is input into the model for forward propagation calculation, and a prediction result is obtained; then, a constructed weighted penalty loss function is used to calculate the error between the model prediction result and the true silicon content, and the network parameters are updated through back propagation; during the training process, the performance of the model is evaluated regularly by using the validation set after image enhancement, and the hyperparameters of the model are adjusted according to the evaluation result; after multiple iterations of training, the network parameters are continuously optimized until a preset number of iterations is reached, at which time the model training is completed, and finally the trained blast furnace molten iron silicon content estimation regression model based on iron sample images is obtained.

4. The iron content estimation method of molten iron in a blast furnace based on an iron-like image according to claim 1 or 3, characterized by, The network structure of the pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images comprises: The first layer is a scaling layer, which scales the image to (1, 224, 224) first, so as to meet the structured operation of the feature extraction network; The second layer is a convolution and pooling layer, which first performs a convolution operation with a kernel size of 7x7, a step of 2 and a padding of 3, and outputs a feature map with a size of (64, 112, 112), wherein 64 represents the channel number of the feature map and 112x112 represents the size of each feature map; secondly, a maximum pooling operation with a pooling window of 3x3, a step of 2 and a padding of 1 is performed, and a feature map with a size of (64, 56, 56) is outputted, wherein 64 represents the channel number of the feature map and 56x56 represents the size of each feature map; The third layer to the tenth layer are residual layers, and the residual layers are composed of convolution, a multi-scale attention module and a residual connection. In one branch, a 3*3 convolution operation is first performed with a step of 1, wherein the step of the fifth layer, the seventh layer and the ninth layer is 2 and the padding is 1, then batch normalization and a ReLU activation function are performed, and then a 3*3 convolution with a step of 1 and padding of 1 is performed, and the output is input into the multi-scale attention module after batch normalization; The multi-scale attention module divides the input features into three branches, two of which are spliced after global pooling in the X and Y directions, and then the channel fusion is performed through a 1*1 convolution to obtain global features; The third branch captures local features through a 3*3 convolution layer. After obtaining two different scale feature representations, corresponding weights are generated, and matrix multiplication calculation is performed with the features of another scale, that is, the weights generated by the features of another scale are used to adjust the features of the current scale, and finally the global spatial attention representation under the current scale is obtained. Finally, the spatial attentions of the two different scales are added, and the weights are generated through a sigmoid function to adjust the input and obtain the final output. In the second branch of the residual layer, the input features are directly kept unchanged. After the features of the two branches are added, the output features are finally output through a ReLU activation function. The eleventh layer performs a global average pooling operation, and the output feature map size is (512, 1, 1); The twelfth layer performs feature calibration, keeps the feature map size unchanged, and finally obtains a feature vector of 512 representing the silicon content of the iron sample; The thirteenth layer uses a fully connected layer to map the (512, 1, 1) feature map to a (1, 1, 1) silicon content value, and finally realizes the estimation task of the silicon content.

5. The iron-like image-based blast furnace hot metal silicon content estimation method according to claim 4, characterized by, In the twelfth layer, a feature calibration module is introduced to smooth the feature distribution between similar silicon content samples, thereby adjusting the extracted feature map.

6. The iron-like image-based blast furnace molten iron Si content estimation method according to claim 5, characterized by, The feature calibration module is used to smooth the feature distribution between similar silicon content samples, and specifically includes: After receiving the features of the previous layer, first, the samples are sorted from small to large according to the silicon content label, and are uniformly divided into multiple intervals; wherein for each silicon content interval, as shown in table b, arranged in ascending order of silicon content, the feature mean μ of the interval is calculated by the following formula b And the standard deviation σ b : In the formula, N b represents the number of samples in the bth silicon content interval, z i is the feature vector of the ith sample in the bth silicon content interval, and represents an element-by-element multiplication operation. Then, the feature information of adjacent intervals is smoothed by a symmetric kernel function, so as to obtain a smoother feature distribution, and the smoothed feature mean value is calculated by the following formula and the standard deviation where k(b, b') is a symmetric kernel function measuring the similarity between the b and b' intervals, B represents the index set of all intervals within the silicon content range; μ b′ represents the characteristic mean of the b' silicon content interval, σ b′ the characteristic standard deviation of the b' silicon content interval; Finally, the feature calibration is performed through standardization and recoloring; z i ′ = (z i - μ b ) / σ b ; where z i represents the normalized eigenvector of the ith sample in the bth silicon content interval, represents the calibration eigenvector of the ith sample in the bth silicon content interval, i = 1,..., N b , b ∈ B.

7. The iron-like image-based blast furnace molten iron Si content estimation method according to claim 3, characterized by, The weighted penalty loss function is: where Loss represents a weighted penalized loss function, where b i is the silicon content interval to which the i-th sample belongs, N represents the total number of samples, y i represents the true silicon content value of the i-th sample, represents the estimated silicon content value of the sample, and a and b are penalty factors.

8. The iron-like image-based blast furnace hot metal silicon content estimation method according to claim 1, characterized by, The evaluation index includes root mean square error RMSE, mean absolute error MAE and cumulative score CS; The calculation formula of the root mean square error RMSE is: The calculation formula of the mean absolute error MAE is: The calculation formula of the cumulative score CS is: wherein represents the actual silicon content of the sample t, represents the predicted silicon content of the model output, N Si represents the total number of samples, H 0.1 (t) and H 0.05 (t) is the Heaviside step function.

9. The iron-like image-based blast furnace hot metal silicon content estimation method according to claim 8, characterized by, The Heaviside step function is defined as follows: where H x (t) is the Heaviside step function, x is the argument.

10. A molten iron silicon content estimation system for a blast furnace based on an iron-like image, characterized by, The system is used to realize the iron sample image-based blast furnace molten iron silicon content estimation method of any one of claims 1 to 9, and specifically includes: An iron sample image collection and calibration module is used to collect iron sample images on the blast furnace site, and calibrate the iron sample images corresponding to each iron sample according to the silicon content of the iron sample; A data set division module is used to construct an iron sample image data set based on the calibrated iron sample images, and divide the iron sample image data set into a training set, a validation set and a test set according to a set proportion; A data enhancement module is used to perform data enhancement processing on the iron sample image data set by using a set data enhancement strategy; The model training module is configured to train a pre-constructed blast furnace molten iron silicon content estimation regression model based on iron sample images by using the image-enhanced training set and the image-enhanced verification set, and obtain a trained blast furnace molten iron silicon content estimation regression model based on iron sample images. The model performance verification module is configured to design an evaluation index and verify the performance of the trained blast furnace molten iron silicon content estimation regression model based on iron sample images by using the image-enhanced test set. The silicon content prediction module is configured to input an iron sample image collected on site into the trained blast furnace molten iron silicon content estimation regression model based on iron sample images, and obtain the silicon content of the iron sample corresponding to the iron sample image.

Images