A blast furnace in-furnace state diagnosis method and system based on ResNet

Abstract

The present application relates to a kind of ResNet-based blast furnace in-furnace state diagnosis method and system.Method includes: S1, constructs blast furnace in-furnace state diagnosis model, including blast furnace in-furnace state classification model, chute state classification model and coal gas flow state classification model;S2, the top imaging image of furnace to be detected is input blast furnace in-furnace state classification model, obtains the probability that this image belongs to normal fluctuation state and the probability of abnormal fluctuation state, and the class of maximum probability is output as classification result;S3, continue the top imaging image of furnace to be detected is input chute state classification model, obtains the probability that this image belongs to no chute, the probability of small chute and the probability of large chute, and the class of maximum probability is output as classification result;S4, continue the top imaging image of furnace to be detected is input coal gas flow state classification model, obtains the probability that this image belongs to center gas flow, the probability of edge gas flow and the probability of two gas flows, and the class of maximum probability is output as last classification result.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of smelting equipment detection, and relates to a blast furnace in-furnace state diagnosis method and system based on ResNet. BACKGROUND

[0002] Blast furnace smelting is a key link in the steel smelting process, and its production cost accounts for more than 60% of the cost of steel products, and energy consumption accounts for about 7% of the energy consumption of the national economy. At the same time, the output of the blast furnace basically determines the production capacity of the steel enterprise. In the face of the common problems of low growth and low efficiency of blast furnace ironmaking, only fundamental upgrading and transformation can continue to develop. With the development of artificial intelligence technology, its application in the industrial field is also constantly innovating. It is an inevitable trend to combine artificial intelligence technology with blast furnace production to improve the long-term safety, stability and efficient operation of blast furnace production.

[0003] The blast furnace has the characteristics of high temperature, high pressure, and high dust. In traditional blast furnace smelting, the internal production of the blast furnace cannot be seen or touched, and the internal information is extremely lacking. At present, there are relatively mature applications of blast furnace monitoring equipment. Among them, blast furnace top imaging is a necessary monitoring technology for monitoring the upper material surface environment of the blast furnace, specifically including video monitoring technology, infrared imaging technology, radar monitoring technology, etc. The above technologies can accurately, intuitively and quickly reconstruct the image information of the material distribution process, to some extent, alleviating the visualization problem of the blast furnace smelting process, and providing a decision basis for blast furnace operation. However, the use of data obtained by the current blast furnace monitoring equipment is obviously insufficient, and the judgment of the blast furnace in-furnace state mainly relies on manual completion. It not only consumes manpower and time, but also is prone to errors. How to collect and organize the monitoring data and combine it with front-end visual perception and other artificial intelligence technologies to realize the diagnosis of the blast furnace in-furnace state is a difficult problem.

[0004] The blast furnace top imaging system is the main equipment for monitoring and detecting the upper material surface environment of the blast furnace. According to different imaging technologies, it mainly includes video monitoring technology, infrared imaging technology, and radar imaging technology. Among them, the video monitoring technology is to use a camera to monitor the material distribution of the top of the furnace; the infrared imaging technology is to use a thermal imager to process the surface infrared image; and the radar imaging technology is to use a radar or distributed radar to monitor the material distribution of the blast furnace. The above technologies can accurately, intuitively and quickly reconstruct the image information of the material distribution process, providing an important data source for the analysis of the blast furnace in-furnace state.

[0005] In the aspect of judging the state in the blast furnace, the most commonly used method is artificial direct observation. The blast furnace operator makes experience judgment by observing the imaging information on the top of the blast furnace, combined with the information of tapping, slagging, tuyere, etc. However, due to the complexity of the state in the blast furnace and the high noise of the photographed image, the artificial observation method highly depends on the experience of the operator and is prone to errors. In addition, some studies propose an indirect observation method based on instrument monitoring. However, this method requires an accurate mathematical model, and the model has high complexity and it is difficult to induce and obtain rules.

[0006] Image classification can be divided into traditional image classification algorithm and deep learning algorithm. The general steps of traditional image classification algorithm are feature extraction, feature coding, feature constraint and classifier classification. With the continuous development of artificial intelligence technology, deep learning algorithm is widely used in image classification task. Convolutional neural network (CNN) is the most common algorithm in image processing task, which has developed rapidly in recent years. It adopts a local connection and weight sharing method to construct a neural network. Image data is input into the neural network for training, and the parameters are adjusted through back propagation. Then the features of the image are extracted through the trained network. The image classification neural network based on CNN includes LeNet-5, AlexNet, VGG, GoogleNet, ResNet, etc. These networks are continuously improved in depth and breadth of convolutional network, which improves the classification effect. SUMMARY

[0007] In view of the above technical problems, one object of the present application is to provide a blast furnace in-furnace state diagnosis method based on ResNet, which can use the blast furnace top image obtained by the blast furnace monitoring device to quickly and accurately automatically identify the in-furnace state, and provide real-time guidance for the comprehensive monitoring, diagnosis and operation optimization of blast furnace production.

[0008] Another object of the present application is to provide a system for implementing the blast furnace in-furnace state diagnosis method based on ResNet, which can realize data collection, data training, diagnosis result display and other functions. The blast furnace technician does not need to master complex technical processing details, and can conveniently operate and view the results.

[0009] In order to achieve the above object, the present application provides the following technical scheme:

[0010] A blast furnace in-furnace state diagnosis method based on ResNet, comprising the following steps:

[0011] S1, constructing a blast furnace in-furnace state diagnosis model, the blast furnace in-furnace state diagnosis model comprising a blast furnace in-furnace state classification model, a chute state classification model and a coal gas flow state classification model; the blast furnace in-furnace state classification model, the chute state classification model and the coal gas flow state classification model are all image classification models based on the ResNet neural network framework;

[0012] S1.1, build a blast furnace in-furnace state classification model;

[0013] A certain number of normal fluctuation state blast furnace top imaging data and a certain number of abnormal fluctuation state blast furnace top imaging data are collected respectively, and a blast furnace top imaging data set is established after data enhancement operation; according to different fluctuation states, the images in the blast furnace top imaging data set are labeled as normal fluctuation state or abnormal fluctuation state, and the labeled data set is randomly divided into a training set and a test set for training and testing the blast furnace in-furnace state classification model respectively;

[0014] S1.2, build a chute state classification model;

[0015] According to different chute conditions, the images labeled as normal fluctuation state in the blast furnace top imaging data set in step S1.1 are labeled as no chute, small chute or large chute; wherein, the image in which the proportion of chute pixels in image pixels is 0 is labeled as no chute, the image in which the proportion of chute pixels in image pixels is greater than 0 and less than or equal to 25% is labeled as small chute, and the image in which the proportion of chute pixels in image pixels is greater than 25% is labeled as large chute; then, the labeled data set is randomly divided into a training set and a test set, which are used for training and testing the chute state classification model respectively;

[0016] S1.3, build a coal gas flow state classification model;

[0017] According to different coal gas flow types, the images labeled as no chute and small chute in the blast furnace top imaging data set in step S1.2 are labeled as central gas flow, edge gas flow or two gas flows; wherein, the image in which the central bright gas flow appears is labeled as central gas flow, the image in which the edge bright gas flow appears is labeled as edge gas flow, and the image in which the central and edge bright gas flows appear is labeled as two gas flows; then, the labeled data set is randomly divided into a training set and a test set, which are used for training and testing the coal gas flow state classification model respectively;

[0018] S2, input the detected top imaging image into the blast furnace in-furnace state classification model, obtain the probability that the image belongs to normal fluctuation state and the probability that the image belongs to abnormal fluctuation state, and output the category with the maximum probability as the classification result; if the output result is abnormal fluctuation state, end the process; if the output result is normal fluctuation state, continue to the next step of chute classification identification;

[0019] S3, continue to input the detected top imaging image into the chute state classification model, obtain the probability that the image belongs to no chute, the probability that the image belongs to small chute and the probability that the image belongs to large chute, and output the category with the maximum probability as the classification result. If the output result is large chute, end the process; if the output result is no chute or small chute, continue to the next step of coal gas flow classification identification;

[0020] S4, continue to input the top of the furnace imaging image to be detected into the gas flow state classification model, obtain the probability that the image belongs to the central gas flow, the probability that the image belongs to the edge gas flow, and the probability that the image belongs to the two gas flows, and output the class with the maximum probability as the final classification result.

[0021] The ResNet neural network introduces a residual block structure, the residual branch of the residual block is built by two layers of convolution, regularization operation and ReLu activation function operation, which is used for nonlinear mapping of input data x to obtain residual representation F(x); the identity mapping connection directly maps x input to output; the output of the final residual block is the sum of the two branch results, that is, F(x)+x; different residual blocks are stacked with each other, and the last layer is a full connection layer, which maps the result to the probability value of each class.

[0022] The residual block structure is Resnet-18 structure, which includes 18 layers with weight parameters, of which 17 are convolution layers and 1 is a full connection layer.

[0023] In the step S1.1, the data enhancement operation includes: rotation, cropping, flipping, scaling, translation and jitter, Gaussian noise, adjusting brightness, histogram equalization, cutout, mixup, and cutmix.

[0024] In the step S1.1, the data enhancement operation uses the transforms method in the python graphics library torchvision, which specifically includes the following steps:

[0025] 1) using transforms.RandomHorizontalFlip(p) to randomly horizontally flip the blast furnace top image obtained by the blast furnace top monitoring device, p is the probability value of the transformation, and a horizontally flipped image is obtained;

[0026] 2) using transforms.RandomVerticalFlip(p) to randomly vertically flip the blast furnace top image obtained by the blast furnace top monitoring device, p is the probability value of the transformation, and a vertically flipped image is obtained;

[0027] 3) using transforms.RandomRotation(p) to randomly rotate the blast furnace top image obtained by the blast furnace top monitoring device, p is the rotation angle range, and a randomly rotated image is obtained;

[0028] 4) using transforms.RandomGrayscale(p) function to convert the blast furnace top image obtained by the blast furnace top monitoring device into a grayscale image according to the probability p.

[0029] A system for implementing the ResNet-based blast furnace in-furnace state diagnosis method comprises a blast furnace in-furnace state diagnosis model training module for training a blast furnace in-furnace state classification model, a chute state classification model and a coal gas flow state classification model;

[0030] A blast furnace in-furnace state diagnosis model storage module is configured to store the blast furnace in-furnace state classification model, the chute state classification model and the coal gas flow state classification model.

[0031] A blast furnace in-furnace state diagnosis result display storage module is configured to display and store the blast furnace in-furnace state diagnosis result.

[0032] The system is built based on winform, the image diagnosis core algorithm is loaded through a python program, and the packaged python program is called through a C# control response function.

[0033] Compared with the prior art, the blast furnace in-furnace state diagnosis method based on ResNet has the following beneficial effects:

[0034] The blast furnace in-furnace state diagnosis method based on ResNet decomposes the blast furnace in-furnace state diagnosis problem into three continuous image classification problems. By using image recognition, visual perception and other artificial intelligence technologies, image data provided by a blast furnace roof monitoring device can be processed to diagnose whether the current blast furnace in-furnace state is normal, whether there is a chute in the blast furnace and the type of the coal gas flow in the blast furnace.

[0035] The system for implementing the ResNet-based blast furnace in-furnace state diagnosis method can realize data collection, data training, diagnosis result display and other functions. Blast furnace technical personnel do not need to master complex technical processing details and can conveniently operate and view the results. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 A blast furnace in-furnace state diagnosis model training flowchart of the present application;

[0037] Figure 2 A neural network learning flowchart of the blast furnace in-furnace state classification model of the present application;

[0038] Figure 3a A blast furnace roof image of an embodiment of the present application;

[0039] Figure 3b A horizontally flipped image of Figure 3a ;

[0040] Figure 3c A vertically flipped image of Figure 3a ;

[0041] Figure 3d A randomly rotated image of Figure 3a ;

[0042] Figure 3e For Figure 3a gray-scale image;

[0043] Figure 4 is a residual block structure schematic diagram;

[0044] Figure 5 is a blast furnace in-furnace state diagnosis flowchart of the present application;

[0045] Figures 6a-6g is the top imaging image of the blast furnace to be detected in the embodiment of the present application. DETAILED DESCRIPTION

[0046] The present application will be further described below in combination with the drawings and embodiments.

[0047] A blast furnace in-furnace state diagnosis method based on ResNet, comprising the following steps:

[0048] S1, constructing a blast furnace in-furnace state diagnosis model, the blast furnace in-furnace state diagnosis model comprising a blast furnace in-furnace state classification model, a chute state classification model and a coal gas flow state classification model;

[0049] As Figure 1 shown, for the characteristics of the blast furnace in-furnace state diagnosis problem, it is first needed to judge whether the blast furnace in-furnace state is a normal fluctuation state; in the normal fluctuation state, the chute state and the gas flow state are further judged in turn. Since the image in the abnormal fluctuation state of the blast furnace will interfere with the subsequent judgment of the coal gas flow, a blast furnace in-furnace state classification model is first constructed for diagnosing whether the blast furnace in-furnace state is normal, the blast furnace in-furnace state classification model taking the imaging data of the top of the blast furnace as input, and the output category being divided into two categories: normal fluctuation state and abnormal fluctuation state; secondly, considering that the chute occupies a large proportion of the image will interfere with the judgment of the coal gas flow, a chute state classification model is constructed for judging the size of the proportion of the chute in the image, the chute state classification model taking the filtered normal fluctuation state data as input, and the output category being divided into three categories: no chute, small chute and large chute; finally, a coal gas flow state classification model is established, taking the filtered data of no chute or small chute as input, and the output being divided into three categories according to the actual blast furnace operation requirements: center gas flow; edge gas flow; two-way gas flow.

[0050] S1.1, constructing a blast furnace in-furnace state classification model; the neural network learning process of the blast furnace in-furnace state classification model is as shown in Figure 2 ;

[0051] A certain amount of blast furnace top imaging data in a normal fluctuation state and a certain amount of blast furnace top imaging data in an abnormal fluctuation state are collected respectively, a blast furnace top imaging data set is established after data enhancement operation, images in the blast furnace top imaging data set are labeled as normal fluctuation state or abnormal fluctuation state according to different fluctuation states, and the labeled data set is randomly divided into a training set and a test set for training and testing a blast furnace internal state classification model respectively.

[0052] Wherein, when a large number of snowflake-shaped points appear in the blast furnace top imaging image or the chute and the internal state of the blast furnace are not clear, it is defined as an abnormal fluctuation state; otherwise, if the above conditions do not exist in the blast furnace top imaging image, it is defined as a normal fluctuation state.

[0053] In order to effectively suppress overfitting, the richness of the training set is improved by more data, and the calculation deviation between the training set and the validation set is reduced. First, the blast furnace top image is preprocessed by data enhancement. The existing image data enhancement technology is applicable to the processing of the present problem. Commonly used data enhancement methods include rotation, cropping, flipping, scaling, translation and jitter, Gaussian noise, brightness adjustment, histogram equalization, cutout, mixup, cutmix, etc.

[0054] The embodiments of the present application take the transforms method in the python graphics library torchvision as an example to introduce the application of several data enhancement strategies in the present problem. It should be noted that other image enhancement technologies, such as Gaussian noise addition, blurring processing, generative adversarial network, style transfer and other image enhancement technologies, can be used.

[0055] The data enhancement operation of the embodiments of the present application specifically includes the following steps:

[0056] 1) Using transforms.RandomHorizontalFlip(p) to randomly horizontally flip the blast furnace top image obtained by the blast furnace top monitoring device (such as shown in Figure 3a ), p is the probability value of transformation, and a horizontally flipped image is obtained, as shown in Figure 3b .

[0057] 2) Using transforms.RandomVerticalFlip(p) to randomly vertically flip the blast furnace top image obtained by the blast furnace top monitoring device (such as shown in Figure 3a ), p is the probability value of transformation, and a vertically flipped image is obtained, as shown in Figure 3c .

[0058] 3) Using transforms.RandomRotation(p) to rotate the blast furnace top image obtained by the blast furnace top monitoring device (such as shown inFigure 3a randomly, p is the angle range of rotation, and a randomly rotated image is obtained, as shown in Figure 3d .

[0059] 4) The blast furnace top image obtained by the blast furnace top monitoring device is converted into a grayscale image according to the probability p using the transforms.RandomGrayscale(p) function, as shown in Figure 3a . Figure 3e

[0060] The blast furnace internal state classification model is an image classification model based on a ResNet neural network framework. In terms of model structure, the core of the ResNet neural network is the introduction of a "residual block" structure. The residual block structure is as shown in Figure 4 . Among them, the residual branch of the residual block is built by two layers of convolution, regularization operation and ReLu activation function operation, which is used to perform nonlinear mapping on the input data x to obtain the residual representation F(x); the identity mapping connection directly maps x to the output; the output of the final residual block is the sum of the two branch results, that is, F(x) + x. Different residual blocks are stacked with each other, and the last layer is a fully connected layer that maps the result to the probability value of each class. Taking the Resnet-18 structure as an example, it contains 18 layers with weight parameters, of which 17 are convolutional layers and 1 is a fully connected layer.

[0061] The training of the blast furnace internal state classification model is an iterative process. As mentioned earlier, the training sample set is the blast furnace top image training sample set after data enhancement preprocessing. In each iteration, the training sample is input, and after forward propagation based on the ResNet deep learning model, the probability that the training sample belongs to the "normal fluctuation state" and the "abnormal fluctuation state" is obtained. Combined with the label of the sample, substitute it into the cross entropy loss function CrossEntropyLoss, and the target deviation can be calculated. Then, the gradient descent method is used for back propagation training. Among them, the Adam algorithm with adaptive learning rate and impulse term is used to realize gradient descent. The reason for choosing the Adam algorithm is that the parameter update not only considers the current gradient value but also adds an impulse term, which can converge faster and better to the point with the smallest error. Secondly, in the Adam algorithm, the learning rate can be decayed based on the accumulated gradient square term, which avoids the rapid gradient descent leading to local optimal solution. Through continuous iteration, the blast furnace internal state classification model parameters (i.e. the weights of the neural nodes in the neural network) can be obtained automatically.

[0062] ​When the model training is completed, the blast furnace in-furnace state classification model can be tested by test samples. After the forward propagation based on the ResNet deep learning model with the test samples as input, the probabilities that the samples belong to the "normal fluctuation state" and the "abnormal fluctuation state" are obtained, and the class with the maximum output probability is taken as the classification result. For example, after a certain test sample is input into the blast furnace in-furnace state classification model, the model outputs the probability of "normal fluctuation state" as 0.6 and the probability of "abnormal fluctuation state" as 0.4, and then the classification result of the sample is "normal fluctuation state".

[0063] S1.2, a chute state classification model is constructed;

[0064] According to different chute conditions, the images in the blast furnace top imaging data set labeled as the normal fluctuation state in step S1.1 are labeled as no chute, small chute or large chute. Among them, the image in which the proportion of chute pixels in image pixels is 0 is labeled as no chute, the image in which the proportion of chute pixels in image pixels is greater than 0 and less than or equal to 25% is labeled as small chute, and the image in which the proportion of chute pixels in image pixels is greater than 25% is labeled as large chute. Then, the labeled data set is randomly divided into a training set and a test set, which are used for training and testing the chute state classification model, respectively.

[0065] Similar to the blast furnace in-furnace state classification model, the chute state classification model is an image classification model based on the ResNet neural network framework. Taking the Resnet-18 structure as an example, it contains 18 layers with weight parameters, of which 17 are convolutional layers and 1 is a fully connected layer. The training of the chute state classification model is an iterative process. In each iteration, after the forward propagation based on the ResNet deep learning model with the training sample as input, the probabilities that the sample belongs to "no chute", "small chute" and "large chute" are obtained. Combined with the label of the sample, the target deviation can be calculated by substituting it into the cross-entropy loss function CrossEntropyLoss. Then, the back propagation training is carried out based on the Adam gradient descent method. Through continuous iteration, the chute state classification model parameters (i.e. the weights on the neural node connection lines in the neural network) can be obtained automatically.

[0066] When the model training is completed, the blast furnace in-furnace state classification model can be tested by test samples. After the forward propagation based on the ResNet deep learning model with the test samples as input, the probabilities that the samples belong to the "normal fluctuation state" and the "abnormal fluctuation state" are obtained, and the class with the maximum output probability is taken as the classification result. For example, after a certain test sample is input into the blast furnace in-furnace state classification model, the model outputs the probability of "normal fluctuation state" as 0.6 and the probability of "abnormal fluctuation state" as 0.4, and then the classification result of the sample is "normal fluctuation state".

[0067] S1.3, build a coal gas flow state classification model;

[0068] According to different types of coal gas flow, the image samples in the blast furnace top imaging data set in step S1.2 are respectively labeled as no chute and small chute, and the image samples are labeled as central gas flow, edge gas flow or two gas flows; wherein, the image samples with bright gas flow appearing in the center are labeled as central gas flow, the image samples with bright gas flow appearing in the edge are labeled as edge gas flow, and the image samples with bright gas flow appearing in the center and edge are labeled as two gas flows; then, the labeled data set is randomly divided into training set and test set two parts, which are respectively used for training and testing the coal gas flow state classification model;

[0069] Similar to the previous two classification models, the coal gas flow state classification model is an image classification model based on the ResNet neural network framework. Taking Resnet-18 structure as an example, it contains 18 layers with weight parameters, of which 17 are convolutional layers and 1 is a fully connected layer. The training of the coal gas flow state classification model is an iterative process, and the training sample set is the image sample set with classification results of "no chute" or "small chute" obtained from S1.2. In each iteration, the training sample is input, and after forward propagation based on the ResNet deep learning model, the probability that the sample belongs to "central gas flow", "edge gas flow" and "two gas flows" is obtained. Combined with the label of the sample, the cross entropy loss function CrossEntropyLoss is substituted into the cross entropy loss function CrossEntropyLoss, and the target deviation can be calculated. Then, based on the Adam gradient descent method, the back propagation training is carried out. Through continuous iteration, the coal gas flow state classification model parameters (i.e. the weights on the neural network nodes) can be obtained automatically.

[0070] When the model training is completed, the coal gas flow state classification model can be tested by test samples. Taking the test sample as input, after forward propagation based on the ResNet deep learning model, the probability that the sample belongs to "central gas flow", "edge gas flow" and "two gas flows" is obtained, and the class with the maximum output probability is output as the classification result. For example, after a certain test sample is input into the coal gas flow state classification model, the model outputs the probabilities of "central gas flow", "edge gas flow" and "two gas flows" as 0.3, 0.5 and 0.2 respectively, and the classification result of the sample is "edge gas flow".

[0071] S2, as Figure 5 shown, input the blast furnace top imaging image to be detected into the blast furnace inner state classification model, obtain the probability that the image belongs to normal fluctuation state and the probability that the image belongs to abnormal fluctuation state, and output the class with the maximum probability as the classification result; if the output result is abnormal fluctuation state, end the process; if the output result is normal fluctuation state, continue to the next step of chute classification and recognition;

[0072] S3, as Figure 5As shown, the furnace top imaging image to be detected is input into the chute state classification model to obtain the probability that the image belongs to no chute, small chute, and large chute, and the category with the highest probability is output as the classification result. If the output result is large chute, the process ends; if the output result is no chute or small chute, the next step of gas flow classification and identification continues.

[0073] S4, such as Figure 5 As shown, the furnace top imaging image to be detected is then input into the gas flow state classification model to obtain the probability that the image belongs to the central gas flow, the probability of the edge gas flow, and the probability of the two gas flows. The category with the highest probability is then output as the final classification result.

[0074] The furnace top imaging image to be detected in this embodiment of the invention is as follows: Figures 6a-6g As shown; where, Figures 6a-6g After inputting the blast furnace internal state classification model, Figures 6b-6g The identification result was "normal fluctuation state", while Figure 6a The identification result was "abnormal fluctuation state". According to... Figure 5 The diagnostic process shown is as follows: Figure 6a This will not proceed to further processing; next, Figures 6b-6g Input the chute state classification model, Figure 6b , Figure 6c , Figure 6f The identification result was "no chute". Figure 6d , Figure 6g The identification result is "small chute". Figure 6e The identification result was "large chute". According to... Figure 5 The diagnostic process shown is as follows: Figure 6e This will not proceed to further processing; finally, Figures 6b-6d , Figure 6f , Figure 6g Input the gas flow state classification model, then Figure 6b The identification result was "central airflow". Figure 6c The identification result was "central airflow". Figure 6d The identification result is "central airflow". Figure 6f The identification result is "central airflow". Figure 6g The identification result was "two airflows".

[0075] Validity verification

[0076] (1) Setting up the dataset

[0077] The total number of samples of the embodiment of the application is 2535, the first classification task is to judge whether the current blast furnace in-furnace state is normal, wherein the number of normal samples is 1427 and the number of abnormal samples is 1108; the second classification task is to judge whether there is a blast furnace in-furnace chute, wherein the number of samples without a chute is 369, the number of small samples with a chute is 411 and the number of large samples with a chute is 647; the third classification task is to judge the airflow type, wherein the number of samples of central airflow is 500, the number of samples of edge airflow is 0 and the number of samples of two-way airflow is 280. The data set is divided into a training set and a test set according to a ratio of 0.85:0.15. Finally, the number of samples of the training set of the first classification task is 2155 and the number of samples of the test set is 380; the number of samples of the training set of the second classification task is 1213 and the number of samples of the test set is 214; and the number of samples of the training set of the third classification task is 663 and the number of samples of the test set is 117.

[0078] (2) Classification evaluation criteria:

[0079] In each classification model, the accuracy rate of the test set classification result is used for model evaluation, that is, the proportion of the number of correct predictions to the total number of the test set, which is expressed by the formula: Wherein, accuracy is the accuracy rate, TP is the number of samples with consistent prediction results and labels, and N is the number of the test set.

[0080] (3) Setting of batch size:

[0081] The batch size refers to the number of samples selected for training at a time. The size of the batch size affects the speed and optimization degree of model training. When the batch size is too large, the training speed is often accelerated, but the GPU memory overhead is large; if the batch size is too small, the training speed is slowed down. In this experiment, the batch size is set to 64.

[0082] (4) Setting of learning rate

[0083] The learning rate is an important hyperparameter in the gradient descent of the neural network. When updating the parameters of the neural network each time, the learning rate determines how much the previous training error is used to update the parameters. The update formula of the neural network parameters is Wherein, α is the learning rate, w is the original neural network parameter, w' is the updated neural network parameter, J(w) is the loss function, The derivative of the loss function with respect to w is represented. The lower the learning rate, the smaller the change of the network parameters, the slower the fitting speed, and the easier the overfitting; the higher the learning rate, the easier the gradient explosion, and the model is difficult to converge. The learning rate selected in the embodiment of the application is 0.001.

[0084] (5) Setting of other parameters

[0085] The ResNet framework includes neural networks of various types of structures, which are all designed based on residual structures and deep networks. An embodiment of the present application selects a Resnet-18 structure, which includes 17 convolutional layers and 1 fully connected layer. The number of iterations (epoch) for each category is 300, 100, and 100, respectively. The neural network optimizer is an Adam optimizer, and the loss function is a cross-entropy loss function CrossEntropyLoss.

[0086] (6) Data augmentation comparison experiment

[0087] The embodiment of the present application uses four commonly used geometric methods for data augmentation, including horizontal flip, vertical flip, random rotation angle, and random grayscale picture. After data augmentation, the number of training set samples for classification task one is 10775, the number of training set samples for classification task two is 6065, and the number of training set samples for classification task three is 3315. The number of test sets does not change. In order to verify the effect of data augmentation, the data set without using data augmentation and the data set using data augmentation are used to train the neural network, and the results are shown in Table 1:

[0088] Table 1 Comparison of image enhancement and non-enhancement experimental results

[0089] Classification Task One Classification Task Two Classification Task Three Without Data Augmentation Method 97.11％ 91.16％ 77.60％ With Data Augmentation Method 97.11％ 93.02％ 88.00％

[0090] From the above table, it can be seen that for classification task one, the effect of the two methods is not much different, but for classification task two and classification task three, the effect of using the data enhancement method is obviously better than that of not using the data enhancement method, which is increased by 1.86% and 10.4%, respectively.

[0091] (7) Comparison experiment of different algorithms

[0092] The advantages of using ResNet are verified by comparing other image classification models. AlexNet, DenseNet, EcaNet, and Vision Transformer methods are selected for comparison with the ResNet model of the present application, and the comparison results are shown in Table 2.

[0093] Table 2 Comparison of ResNet and other model experimental results

[0094] Classification Task One Classification Task Two Classification Task Three AlexNet 94.23％ 88.37％ 81.60％ DenseNet 96.59％ 90.23％ 77.60％ EcaNet 96.85％ 91.63％ 81.60％ Vision Transformer 67.45％ 62.33％ 74.40％ ResNet 97.11％ 93.02％ 88.00％

[0095] From the above table, it can be seen that the ResNet model is the best in the three classifications, which shows its strong robustness in industrial applications. Vision Transformer performs poorly in the three classifications, which shows that the small number of data sets is not conducive to training the network.

[0096] In summary, by enhancing the image data, using the ResNet network model to train the image data set of the blast furnace in the state, and finally the test results on the test set show that the selected ResNet is the best.

[0097] The application further provides a system for implementing the above-mentioned ResNet-based blast furnace in-furnace state diagnosis method, and the system comprises:

[0098] a blast furnace in-furnace state diagnosis model training module for training a blast furnace in-furnace state classification model, a chute state classification model and a coal gas flow state classification model;

[0099] a blast furnace in-furnace state diagnosis model storage module for storing the blast furnace in-furnace state classification model, the chute state classification model and the coal gas flow state classification model;

[0100] a blast furnace in-furnace state diagnosis result display and storage module for displaying and storing the blast furnace in-furnace state diagnosis result.

[0101] The system is built based on winform and mainly comprises three functions: single image diagnosis, image batch diagnosis and export of prediction results. The core algorithm of image diagnosis is loaded through a python program, and the packaged python program is called through a C# control response function.

[0102] The C# window form program winform can be used as a platform for various controls, has rich user resources, can develop and design corresponding clients, and can be used in networking and single machine conditions, and can safely access resources on the local computer. Considering that the ResNet deep learning model is implemented based on the pytorch framework, the saved model also needs to be loaded through a python program. Therefore, the python program is embedded into the winform window form program after being packaged.

[0103] The specific implementation process of the winform window form program is as follows: the blast furnace top imaging data set is input into the neural network for training, and the best-performing model is saved; the saved model is called through a python program; the python program is packaged into an.exe file using the python packaging instruction pyinstaller; a new winform program is created, the interface is designed, and the controls are arranged; the packaged.exe file is called in the response function of the control using an interface function; the.exe file is parsed and the response result is returned; finally, the result is filled into the winform interface for visual display. In order to meet the needs of in-furnace diagnosis, the winform program comprises the following functions:

[0104] (1) Single image diagnosis

[0105] In the interface, there is a button of "select image". By clicking this button, the system can find the single image of blast furnace top from local files, and display the image to PictureBox. By clicking the button of "predict", the system can automatically call the packaged Python program to classify, and display the classification results of the three categories and the time consumed in the prediction to RichTextBox.

[0106] (2) Batch diagnosis of images

[0107] In the interface of diagnosing multiple images of blast furnace, there is a button of "select folder". By clicking this button, the user can select the folder address where the images are stored. By clicking the button of "predict", the system can call the program for classifying multiple images, and finally return all the classification results and display them to RichTextBox.

[0108] (3) Export results

[0109] The results of image diagnosis can be saved to a form file. In the interface of exporting prediction results, the user can click the button of "select the folder to be exported" to select the path to be exported. Then the system automatically reads the path and displays it to Label. Finally, by clicking the button of "export", the system can export all the results to the specified folder.

Claims

1. A method for diagnosing the internal condition of a blast furnace based on ResNet, characterized in that, The method includes the following steps: S1. Construct a blast furnace internal state diagnostic model, which includes a blast furnace internal state classification model, a chute state classification model, and a gas flow state classification model; the blast furnace internal state classification model, the chute state classification model, and the gas flow state classification model are all image classification models based on the ResNet neural network framework; S1.1 Construct a classification model for the internal state of a blast furnace; A certain number of blast furnace top imaging data under normal fluctuation conditions and a certain number of blast furnace top imaging data under abnormal fluctuation conditions were collected respectively. After data augmentation, a blast furnace top imaging dataset was established. According to different fluctuation conditions, the images in the blast furnace top imaging dataset were labeled as normal fluctuation conditions or abnormal fluctuation conditions. The labeled dataset was randomly divided into a training set and a test set, which were used to train and test the blast furnace internal state classification model, respectively. S1.2 Construct a chute state classification model; Based on different chute conditions, images in the blast furnace top imaging dataset from step S1.1 labeled as being in a normal fluctuation state are marked as having no chute, having a small chute, or having a large chute. Specifically, images where the proportion of chute pixels to image pixels is 0 are marked as having no chute, images where the proportion of chute pixels to image pixels is greater than 0 and less than or equal to 25% are marked as having a small chute, and images where the proportion of chute pixels to image pixels is greater than 25% are marked as having a large chute. Then, the labeled dataset is randomly divided into a training set and a test set, which are used to train and test the chute state classification model, respectively. S1.3 Construct a coal gas flow state classification model; Based on different gas flow types, the images labeled as having no chute and having a small chute in the blast furnace top imaging dataset in step S1.2 are labeled as central gas flow, edge gas flow, or two gas flow; among them, images with a bright gas flow in the center are labeled as central gas flow, images with a bright gas flow at the edge are labeled as edge gas flow, and images with a bright gas flow in both the center and the edge are labeled as two gas flow; then, the labeled dataset is randomly divided into a training set and a test set, which are used to train and test the gas flow state classification model, respectively. S2. Input the furnace top imaging image to be detected into the blast furnace internal state classification model to obtain the probability that the image belongs to the normal fluctuation state and the probability of the abnormal fluctuation state, and output the category with the highest probability as the classification result; if the output result is the abnormal fluctuation state, the process ends; if the output result is the normal fluctuation state, continue to the next step of chute classification and recognition. S3. Continue to input the furnace top imaging image to be detected into the chute state classification model to obtain the probability that the image belongs to no chute, small chute, and large chute, and output the category with the highest probability as the classification result; if the output result is large chute, the process ends; if the output result is no chute or small chute, continue to the next step of gas flow classification and identification. S4. Continue to input the furnace top imaging image to be detected into the gas flow state classification model to obtain the probability that the image belongs to the central gas flow, the probability of the edge gas flow, and the probability of the two gas flows, and output the category with the highest probability as the final classification result.

2. The blast furnace in-furnace condition diagnosis method based on ResNet according to claim 1, characterized in that, The ResNet neural network introduces a residual block structure. The residual branch of the residual block is constructed by two layers of convolution, regularization operations, and ReLU activation function operations to perform nonlinear mapping on the input data x to obtain the residual representation F(x). The identity mapping connection directly maps the input x to the output. Finally, the output of the residual block is the sum of the results of the two branches, which is F(x) + x. Different residual blocks are stacked on top of each other, and the last layer is a fully connected layer that maps the result to the probability value of each class.

3. The blast furnace in-furnace condition diagnosis method based on ResNet according to claim 2, characterized in that, The residual block structure is a ResNet-18 structure, which contains 18 layers with weight parameters, including 17 convolutional layers and 1 fully connected layer.

4. The blast furnace internal condition diagnosis method based on ResNet according to claim 1, characterized in that, In step S1.1, the data augmentation operations include: rotation, cropping, flipping, scaling, translation and dithering, Gaussian noise, brightness adjustment, histogram equalization, cutout, mixup, and cutmix.

5. The blast furnace in-furnace condition diagnosis method based on ResNet according to claim 1, characterized in that, In step S1.1, the data augmentation operation uses the transforms method from the Python graphics library torchvision, specifically including the following steps: 1) Use transforms.RandomHorizontalFlip(p) to randomly flip the blast furnace top image obtained by the blast furnace top monitoring equipment horizontally, where p is the probability value of the transformation, to obtain a horizontally flipped image; 2) Use transforms.RandomVerticalFlip(p) to randomly vertically flip the blast furnace top image obtained by the blast furnace top monitoring equipment, where p is the probability value of the transformation, to obtain a vertically flipped image; 3) Use transforms.RandomRotation(p) to randomly rotate the blast furnace top image obtained by the blast furnace top monitoring equipment, where p is the range of rotation angles, to obtain a randomly rotated image; 4) Use the transforms.RandomGrayscale(p) function to convert the blast furnace top image obtained by the blast furnace top monitoring equipment into a grayscale image based on probability p.

6. A system for implementing the ResNet-based blast furnace in-furnace condition diagnosis method as described in any one of claims 1-5, characterized in that, The system includes a blast furnace in-furnace condition diagnosis model training module, which is used to train a blast furnace in-furnace condition classification model, a chute condition classification model, and a gas flow condition classification model. The blast furnace in-furnace condition diagnosis model storage module is used to store the blast furnace in-furnace condition classification model, the chute condition classification model, and the gas flow condition classification model. The blast furnace in-furnace condition diagnosis result display and storage module is used to display and store the blast furnace in-furnace condition diagnosis results.

7. The system according to claim 6, characterized in that, The system is built on WinForms. The core image diagnosis algorithm is loaded through a Python program and then implemented by calling the packaged Python program through C# control response functions.

Images