A Blast Furnace Fault Diagnosis Method Based on Minimax Entropy Co-Training

Abstract

The present invention discloses a blast furnace fault diagnosis method based on minimax entropy collaborative training, belonging to the fields of industrial process monitoring, modeling and simulation. First, a dual-view feature extractor is constructed using a deep neural network to extract features from the historical data and the data to be measured of the blast furnace, and its outputs are respectively fed into two classifiers to calculate the cosine similarity between the features and the representative vectors of each furnace condition. The knowledge transfer of the historical blast furnace data is realized by alternately maximizing and minimizing the conditional entropy of the blast furnace data to be measured by the classifier and the feature extractor respectively. When the recognition results of the two classifiers for the blast furnace condition are consistent and at least one classifier has a high confidence level, the output furnace condition label is assigned to the blast furnace data to be measured. The present invention solves the problem of low accuracy of traditional fault diagnosis methods caused by few blast furnace fault samples and large fluctuations in data distribution with the change of working conditions, and can be widely applied to industrial systems with high requirements for the credibility and accuracy of fault diagnosis.

Description

Technical Field

[0001] The present invention belongs to the fields of industrial process monitoring, modeling and simulation, and particularly relates to a blast furnace fault diagnosis method based on minimax entropy co-training. Background Art

[0002] Blast furnace ironmaking is the core unit of the conversion of ferrous material flow and is the link with the highest energy consumption and production cost in the steel manufacturing process. With the continuous progress of the process and technology in the blast furnace ironmaking process and the continuous development of instrumentation and automation technology, modern blast furnace ironmaking shows characteristics such as large scale, complex structure, strong coupling between production units, and huge investment. If abnormal fluctuations (or accidents) in the blast furnace ironmaking process are not detected in time, it often leads to a serious decline in product quality, or delays the normal execution of the production plan, causing huge economic losses and casualties. Blast furnace fault diagnosis is of great significance for ensuring the safe and efficient production of blast furnaces.

[0003] In the actual production process of blast furnace ironmaking, in order to avoid serious consequences, when certain fault omens occur in the operation of the blast furnace system, operators will adjust the air supply system, burden distribution system or furnace heat system to avoid the occurrence of faults. Therefore, under the existing operating system and operating conditions, building a blast furnace fault diagnosis system faces problems such as few fault samples, data imbalance, missing labels, and high cost and time consumption for labeled samples. In addition, due to the problem of non-fixed raw material origins, most of the blast furnace ironmaking feed in domestic steel mills adopts the form of "hundred mines". At different times, the types and their ratios of the feed will change significantly. Secondly, there are multiple operating condition switches in the production operation process of the blast furnace. These factors all cause the blast furnace data to change over time, with large fluctuations in data distribution, and there are distribution differences between the training data and the data to be measured, affecting the reliability and accuracy of fault diagnosis.

[0004] Currently, the fault diagnosis methods applied to blast furnaces can be roughly divided into two types, namely expert systems and data-driven intelligent fault diagnosis methods. Expert systems have high requirements for prior knowledge such as relevant knowledge and rules, and the physical and chemical reactions involved in blast furnaces are extremely complex, and it is difficult to know the accurate situation of the actual reactions occurring inside. Moreover, with the widespread use of various intelligent instruments and control devices in distributed control systems in modern industrial processes, a large amount of process data has been collected and stored. However, this data containing process operation status information is often not effectively utilized in expert systems.

[0005] On the other hand, the successful application of traditional data-driven intelligent fault diagnosis methods has two prerequisites: 1) a large amount of labeled data 2) training and test data come from the same data distribution. However, in the blast furnace production process, labeled fault samples are very few and difficult to obtain. In addition, due to the difference in the grade of ore raw materials and production conditions, the data will fluctuate greatly, resulting in the training data and test data often not meeting the same distribution conditions. Therefore, the existing abnormal furnace condition diagnosis methods are still far from practical application, and new paths and methods need to be explored. Summary of the invention

[0006] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a blast furnace fault diagnosis method based on minimal maximum entropy collaborative training. The method first uses a deep neural network to construct a dual-view feature extractor to extract features from the historical data and the data to be tested of the blast furnace, and then its output is respectively sent to two classifiers to calculate the cosine similarity between the feature and the representative vector of each furnace condition. The knowledge transfer of historical blast furnace data is realized by alternately maximizing and minimizing the conditional entropy of the blast furnace data to be tested by the classifier and the feature extractor. When the two classifiers have the same recognition results for the blast furnace condition and at least one classifier has a high confidence, the output blast furnace condition is assigned to the blast furnace data to be tested. Experiments using actual industrial blast furnace data show that the method has achieved good results in diagnosing abnormal conditions in blast furnace ironmaking. The method of the present invention not only improves the diagnostic accuracy by using a deep neural network, but also solves the problem of low accuracy of the traditional fault diagnosis method caused by the small number of blast furnace fault samples and the large fluctuation of data distribution with changes in working conditions by using the minimal maximum entropy method and the collaborative training method, and can be widely used in industrial systems with high credibility and accuracy requirements for fault diagnosis.

[0007] A blast furnace fault diagnosis method based on minimax entropy collaborative training, the steps are as follows:

[0008] Step 1: Use the blast furnace historical data to train the weights of the deep neural network. The vector value in the last fully connected layer in the feature extractor of the neural network is the extracted feature value. The L1 norm sum of the fault diagnosis error of the blast furnace historical data and the weighted inner product of the dual views is used as the loss function. After the training reaches a preset number of iterations or the loss function is lower than a preset value, the weight is fixed;

[0009] Step 2: The neural network forms a nonlinear mapping from blast furnace process variables to blast furnace fault categories, classifies the labeled blast furnace data through the neural network, and calculates the classification loss value;

[0010] Step 3: Input the unlabeled blast furnace into the neural network and calculate the entropy value of the neural network output value;

[0011] Step 4: Use the stochastic gradient descent method to update the classifier parameters in the neural network with the goal of minimizing the classification loss value and maximizing the entropy value, so that the weight vector of the classifier shifts towards the unlabeled blast furnace data to be measured, and obtain the class center in the blast furnace data to be measured;

[0012] Step 5: The feature vectors of the blast furnace data to be measured converge towards the corresponding class centers to achieve low-density separation of the blast furnace data to be measured. Use the stochastic gradient descent method to update the feature extractor parameters in the neural network with the goal of minimizing the classification loss value and minimizing the entropy value;

[0013] Step 6: When the blast furnace condition recognition results of the two classifiers are consistent and at least one classifier has a high confidence in this result, assign the output blast furnace condition label value to the blast furnace data to be measured;

[0014] Step 7: Remove the blast furnace data to be measured with the obtained condition label value from the unlabeled blast furnace dataset and add it to the labeled blast furnace dataset;

[0015] Step 8: Perform cyclic iteration on Steps 2 to 7 until the condition recognition of all unlabeled blast furnace data to be measured is completed.

[0016] The structure of the deep neural network described in Step 1 is as follows: The deep neural network consists of an input layer, a hidden layer, and a dual-view output layer; The feature extractor includes an input layer and a hidden layer. The input layer is the input layer of blast furnace process variable parameters, and the parameters are industrial process parameters characterizing the production status of the blast furnace, including the permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure. A blast furnace sample consists of a matrix composed of blast furnace process variable parameters at 35 moments; The dual-view output layer is the fault category layer related to the blast furnace production process, and the outputs include stickiness, stock hanging, channeling, collapse, furnace heat, and furnace coolness; The hidden layer is used to establish a non-linear mapping from blast furnace process variables to blast furnace fault categories, so as to learn blast furnace fault diagnosis knowledge from blast furnace historical fault data and establish a blast furnace fault diagnosis model; Neurons in the same layer are not connected, and neurons between layers are fully connected. Each connection has a weight value, which represents the strength of the connection between neurons; The output features of the last fully connected layer of the sample in the feature extractor pass through different weight matrices W 1 and W 2 are divided into two mutually exclusive views to meet the conditions of co-training;

[0017] (1)

[0018] where d is the dimension of the weight vector. For the convenience of parameter optimization, use the L1 norm of the inner product of the weight matrix To approximately replace this constraint, the dual view is specifically manifested as two different fully connected layers in the neural network, and their outputs are sent to two classifiers for blast furnace condition recognition respectively; perform a normalization operation on the feature vectors of the dual view. When the prediction results of classifiers C1 and C2 are consistent and at least one classifier has a high confidence in the prediction result, assign pseudo-labels to the unlabeled target samples , the model needs to correctly classify the labeled blast furnace data under the condition that the input features of the two classifiers are different. The loss function for pre-training the network is defined as minimizing the classification loss and L1 norm of the labeled data ; The objective function is defined as follows

[0019] (2)

[0020] is the standard cross-entropy loss function D l is the labeled blast furnace data set, (x, y ) is the blast furnace sample and the corresponding condition label value represents the expected value function y is the blast furnace data condition label value p ( x ) is the recognition result of the blast furnace sample condition by the output layer of the neural network

[0021] The steps to update the classifier parameters described in step four are as follows: Entropy is introduced into the model to learn the features with classification information in the blast furnace data to be measured. The entropy calculation process is as follows

[0022] (3)

[0023] where D u represents the unlabeled blast furnace data set y is the blast furnace sample corresponding condition label value output by the classifier represents the expected value function is the blast furnace sample x in the output layer belongs to the i th class of blast furnace condition probability value k represents the number of blast furnace condition categories

[0024] Increase the similarity between the classifier weight W and the features of the blast furnace data to be measured by maximizing the entropy, so that the weight vector W of the linear classifier approaches the distribution of the blast furnace data to be measured and obtain the class center in the blast furnace data to be measured

[0025] (4)

[0026] represent the weight parameters of the classifier, is its corresponding optimal solution, is a hyperparameter to adjust the influence of entropy value and classification error.

[0027] The steps for updating the feature extractor parameters described in Step Five are as follows: Generate the required features with classification information by reducing the classification loss and entropy value of the feature extractor,

[0028] (5)

[0029] represent the weight parameters of the feature extractor, is its corresponding optimal solution, is a hyperparameter to adjust the influence of entropy value and classification error.

[0030] The iterative loop described in Step Eight: Adversarial learning between the classifier and the feature extractor. While reducing the classification loss, perform maximum entropy training on the classifier and minimum entropy training on the feature extractor. This process is continuously iterated to ensure the performance of the model.

[0031] The method described is applied to the blast furnace ironmaking production process with multiple working condition switches, variable feed types, unstable ore quality, large fluctuations in data distribution, and few data labels.

[0032] Advantages of the present invention:

[0033] Aiming at the characteristics and scientific problems of large data fluctuations, missing labels, and data imbalance in the blast furnace ironmaking process, a blast furnace fault diagnosis method based on min-max entropy collaborative training is constructed, fully mining the knowledge contained in blast furnace data, solving the problem that although there is a large amount of blast furnace historical data, it is difficult to directly train a model for the data to be measured. At the same time, it has the advantages of high reliability and high accuracy, improving the automation and intelligent level of the ironmaking process. Brief Description of the Drawings

[0034] Figure 1 The flowchart of the method of the present invention is shown.

[0035] Figure 2 The visualization result of the original distribution of the data to be measured by t-sne is shown.

[0036] Figure 3 The visualization result of the blast furnace fault classification of the blast furnace data to be measured by the method of the present invention by t-sne is shown. Detailed Embodiment

[0037] The purpose of the present invention is to provide a blast furnace fault diagnosis method based on a weighted joint distribution adaptation neural network, and the flowchart is asFigure 1 As shown, considering the non - linear and non - Gaussian properties of blast furnace production information, the advantage that a deep neural network can infinitely approximate a non - linear function is utilized for feature extraction. The model structure includes a two - view feature extractor, and then two classifiers are correspondingly set up to calculate the cosine similarity between the features and the center of each class. When the recognition results of the blast furnace condition by the two classifiers are consistent and at least one classifier has a high confidence level, the output blast furnace condition is assigned to the blast furnace data to be measured. By maximizing entropy, the similarity between the classifier weights and the features of the blast furnace data to be measured is increased, so that the weight vector of the linear classifier approaches the distribution of the blast furnace data to be measured. By reducing the classification loss and entropy value of the feature extractor, the features of the blast furnace data to be measured are aggregated towards the class center to generate features with classification information on the blast furnace data to be measured. Finally, through an iterative cycle, the performance of the model in blast furnace fault diagnosis is ensured. This method helps to achieve knowledge and decision - making enhancement in blast furnace fault diagnosis and guarantees the credibility and accuracy of blast furnace fault diagnosis. Next, the fault data of the No. 2 blast furnace collected from a steel plant is used to verify the effectiveness of the method of the present invention.

[0038] The blast furnace is divided into five parts from top to bottom: the throat, the shaft, the bosh, the hearth and the hearth bottom. Coke, ore and flux will experience different changes in different parts of the furnace during the descending process until they are completely converted into molten iron and slag at the bottom of the hearth. Due to the huge volume of the blast furnace and the complex chemical reactions occurring inside, ensuring its safe and stable operation is extremely important. Blast furnace faults are mainly divided into 4 categories: difficult, hanging, channeling, and collapsing. The data collected during the production process includes 29 parameters such as the permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure. In actual production, a three - shift system is adopted to organize workers to monitor and operate the blast furnace iron - making process, which consumes a large amount of labor costs, and the control method is relatively extensive. The furnace condition is judged mainly based on several parameters, and it is difficult to diagnose the problems existing in the blast furnace operation process in a timely manner and carry out precise control in a timely manner. The method of the present invention can solve this problem to a certain extent and has practical application value.

[0039] Next, the implementation steps of the present invention will be elaborated in detail in combination with this specific process:

[0040] I. Construct a deep neural network for feature extraction

[0041] (1) The deep neural network consists of three parts: an input layer, a hidden layer, and a dual-view output layer. The feature extractor includes two parts: an input layer and a hidden layer. The input layer is the input layer of blast furnace process variables, including industrial process parameters such as permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure, which characterize the production status of the blast furnace. Considering the time series nature of blast furnace data, we extended the input blast furnace samples. A blast furnace sample consists of a matrix of blast furnace process variable parameters at 35 moments. The dual-view output layer is the blast furnace fault category layer, and the outputs include blast furnace faults related to the blast furnace production process, such as stockline creep, hanging, channeling, collapse, heat up, and cool down. The role of the hidden layer is to establish a non-linear mapping from blast furnace process variables to blast furnace fault categories. Therefore, blast furnace fault diagnosis knowledge can be learned from blast furnace historical fault data, and a blast furnace fault diagnosis model can be established. Neurons in the same layer are not connected, and neurons between layers are fully connected. Each connection has a weight value, which characterizes the strength of the connection between neurons. For different industrial application fields, the requirements for the number of hidden layers of the deep neural network are different. A neural network with a hidden layer greater than or equal to 2 is defined as a deep neural network. The mathematical model of the deep neural network is:

[0042]

[0043] Among them, is the output of the j-th hidden layer unit in the i-th layer of the neural network, denoted as is the i-th layer of the neural network, then h 0 is the input layer of the neural network, h k+1 is the output layer of the neural network; the value of j is determined according to the number of neurons in the i-th layer of the network. Denote the number of neurons in the i-th layer as Z i , then the value of j for each layer ranges from 1 to Z i ; is the weight matrix corresponding to the j-th neuron in the i-th layer; is the bias term corresponding to the j-th neuron in the i-th layer, b k+1 is the bias term corresponding to the output layer unit, and are the activation functions of the hidden layer unit and the output unit respectively, and y represents the output of the neural network.

[0044] (2) We divide the output features of the last fully connected layer of the sample in the feature extractor into two mutually exclusive views through different weight matrices W 1 and W 2 to meet the conditions of co-training.

[0045] (1)

[0046] Where d is the dimension of the weight vector. For ease of parameter optimization, we use the L1 norm of the inner product of the weight matrices to approximately replace this constraint. In the neural network, the dual-view is specifically manifested as two different fully connected layers, and their outputs are sent to two classifiers respectively for blast furnace condition recognition. Since the direction of the weight vector of the classifier can often represent the normalized features of the relevant categories, we perform a normalization operation on the feature vectors of the dual-view. Taking classifier C1 as an example, its weight vector is , where k represents the total number of blast furnace conditions. C1 takes the normalized feature vector as the input, and the output is , where is the output value of the corresponding view of C1. Then, the output of C1 is fed into the softmax layer to obtain the probability distribution of various blast furnace conditions, which we denote as

[0047] (8)

[0048] represents the softmax function, is the blast furnace data, and the working process of classifier C2 is basically the same as that of C1, which will not be repeated here. Finally, when the prediction results of C1 and C2 are consistent, and when at least one classifier has a high confidence in the prediction result, we assign a pseudo-label to the unlabeled target sample.

[0049] The model needs to correctly classify the labeled blast furnace data under the condition that the input features of the two classifiers are different. Therefore, the loss function for pre-training the network can be defined as minimizing the classification loss of the labeled data and the L1 norm . The objective function can be defined as follows.

[0050] (2)

[0051] is the standard cross-entropy loss function, D l is the labeled blast furnace data set, (x, y ) is the blast furnace sample and the corresponding condition label value, represents the expected value function, y is the blast furnace data condition label value, p ( x ) is the recognition result of the blast furnace sample condition by the output layer of the neural network.

[0052] II. Maximum Entropy Training of the Classifier

[0053] Since there is no data of the blast furnace to be measured in the initial labeled blast furnace dataset, the weight vector of the classifier will deviate from the distribution of the blast furnace data to be measured. To learn the features with classification information in the blast furnace data to be measured, we introduce the entropy value into the model.

[0054] (3)

[0055] where D u represents the unlabeled blast furnace dataset, y is the corresponding furnace condition label value of the blast furnace sample output by the classifier; represents the expected value function, is the blast furnace sample x belongs to the i th class of furnace condition in the output layer, k represents the number of blast furnace condition categories.

[0056] To obtain the class center in the blast furnace data to be measured, we increase the similarity between the classifier weight W and the features of the blast furnace data to be measured by maximizing the entropy, so that the weight vector W of the linear classifier approaches the distribution of the blast furnace data to be measured. To obtain the class center in the blast furnace data to be measured.

[0057] (4)

[0058] represents the weight parameter of the classifier, is its corresponding optimal solution, is a hyperparameter to adjust the influence of entropy value and classification error.

[0059] III. Minimize the entropy training of the feature extractor

[0060] To obtain the features with classification information in the blast furnace data to be measured, we need to cluster the features of the blast furnace data to be measured at the class center. We generate the required features with classification information by reducing the classification loss and entropy value of the feature extractor.

[0061] (5)

[0062] represents the weight parameter of the feature extractor, is its corresponding optimal solution, is a hyperparameter to adjust the influence of entropy value and classification error.

[0063] IV. Perform iterative solution

[0064] By iteratively looping steps two and three, our method can be formulated as adversarial learning between a classifier and a feature extractor. While reducing the classification loss simultaneously, the classifier is trained with maximum entropy, and the feature extractor is trained with minimum entropy. To achieve adversarial learning, we continuously iterate this process to ensure model performance.

[0065] V. Substitute industrial actual data for verification

[0066] We take the production data of No. 2 blast furnace with a volume of 2650m 3 in a certain ironmaking plant in October 2017 as the blast furnace historical data, i.e., the source domain data, and the blast furnace production data in November as the data to be measured, i.e., the target domain data, which contains 35 parameters and has a consistent sampling rate. The model obtained from the training is verified for effectiveness on the data to be measured, i.e., the blast furnace production data in November.

[0067] Figure 2 , as shown in Figure 3, are respectively the original distribution of the blast furnace data to be measured in November and the visualization result after the blast furnace fault classification of the data to be measured by the method of the present invention using t-sne. It can be seen from the fault diagnosis results that the model effect is very good. The classification effect is obvious, and it can accurately classify the blast furnace fault samples. Therefore, it can be applied to actual industrial production.

[0068] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the appended claims.

Claims

1. A blast furnace fault diagnosis method based on minimax entropy co-training, characterized in that The steps are as follows: Step 1: Use the historical data of the blast furnace to train the weights of the deep neural network. The vector value in the last fully connected layer of the feature extractor in the neural network is the extracted feature value. Take the sum of the L1 norms of the inner products of the fault diagnosis errors of the blast furnace historical data and the weights of the dual views as the loss function, and fix the weights after the training reaches the preset number of iterations or the loss function is lower than the preset value; Step 2: The neural network forms a non-linear mapping from the blast furnace process variables to the blast furnace fault categories. Classify the labeled blast furnace data through the neural network calculation, and calculate the classification loss value; Step 3: Input the unlabeled blast furnace into the neural network and calculate the entropy value of the output value of the neural network; Step 4: Use the stochastic gradient descent method to update the classifier parameters in the neural network with the goal of minimizing the classification loss value and maximizing the entropy value, so that the weight vector of the classifier shifts towards the unlabeled blast furnace data to be measured, and obtain the class center in the blast furnace data to be measured; Step 5: The feature vectors of the blast furnace data to be measured gather towards the corresponding class centers to achieve the low-density separation of the blast furnace data to be measured. Use the stochastic gradient descent method to update the feature extractor parameters in the neural network with the goal of minimizing the classification loss value and minimizing the entropy value; Step 6: When the recognition results of the blast furnace condition by the two classifiers are consistent and at least one classifier has a high confidence in this result, assign the output blast furnace condition label value to the blast furnace data to be measured; Step 7: Remove the blast furnace data to be measured with the obtained blast furnace condition label value from the unlabeled blast furnace dataset and add it to the labeled blast furnace dataset; Step 8: Perform cyclic iteration on Steps 2 to 7 until the blast furnace condition recognition of all unlabeled blast furnace data to be measured is completed.

2. The method according to claim 1, wherein The structure of the deep neural network described in Step 1 is as follows: The deep neural network includes an input layer, a hidden layer, and a dual-view output layer; the feature extractor includes an input layer and a hidden layer. The input layer is the input layer of the blast furnace process variable parameters, and the parameters are industrial process parameters representing the production state of the blast furnace, including the permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure. A blast furnace sample is composed of a matrix of blast furnace process variable parameters at 35 moments; the dual-view output layer is the fault category layer related to the blast furnace production process, and the outputs include difficult operation, stock hanging, channeling, collapse, furnace heat, and furnace cool; the hidden layer is used to establish a non-linear mapping from the blast furnace process variables to the blast furnace fault categories, so as to learn the blast furnace fault diagnosis knowledge from the blast furnace historical fault data and establish a blast furnace fault diagnosis model; the neurons in the same layer are not connected, and the neurons between layers are fully connected, and each connection has a weight value, which represents the strength of the connection between neurons; Pass the output features of the last fully connected layer of the sample in the feature extractor through different weight matrices W 1 and W 2 divide them into two mutually exclusive views to meet the conditions for co-training; (1) Where d is the dimension of the weight vector. To facilitate parameter optimization, the L1 norm of the inner product of the weight matrices is used to approximately replace this constraint. In the neural network, the dual view is specifically manifested as two different fully connected layers, and their outputs are sent to two classifiers respectively for blast furnace condition identification; the feature vectors of the dual view are normalized. When the prediction results of classifiers C1 and C2 are consistent, and when at least one classifier has a high confidence in the prediction results, pseudo-labels are assigned to the unlabeled target samples , the model needs to correctly classify the labeled blast furnace data under the condition that the input features of the two classifiers are different. The loss function for pre-training the network is defined as minimizing the classification loss and the L1 norm of the labeled data ; The objective function is defined as follows, (2) is the standard cross-entropy loss function, D l is the labeled blast furnace dataset, (x, y ) is the blast furnace sample and the corresponding furnace condition label value, represents the expected value function, y is the furnace condition label value of the blast furnace data, p ( x ) is the recognition result of the neural network output layer for the furnace condition of the blast furnace sample.

3. The method according to claim 2, wherein The steps for updating the classifier parameters described in Step 4 are as follows: Entropy value is introduced into the model to learn the features with classification information in the blast furnace data to be measured. The entropy value calculation process is as follows: (3) Among them D u represents the tagless blast furnace dataset, y which is the blast furnace sample's corresponding furnace condition tag value output by the classifier; represents the function for calculating the expected value, is the blast furnace sample x in the output layer and belongs to the i probability value of the furnace condition of the k category, representing the number of blast furnace condition categories; Increase the similarity between the classifier weight W and the features of the blast furnace data to be measured by maximizing entropy, so that the weight vector W of the linear classifier approaches the data distribution of the blast furnace to be measured, and the class center in the blast furnace data to be measured is obtained. (4) represents the weight parameters of the classifier, which is its corresponding optimal solution, and is a hyperparameter to adjust the influence of entropy value and classification error.

4. The method according to claim 3, characterized in that The steps for updating the feature extractor parameters described in Step 5 are as follows: Generate the required features with classification information by reducing the classification loss and entropy value of the feature extractor. (5) representing the weight parameters of the feature extractor, which is its corresponding optimal solution, and is a hyperparameter to adjust the influence of entropy value and classification error.

5. The method according to claim 1, characterized in that, The iterative loop described in Step 8: Adversarial learning between the classifier and the feature extractor. While reducing the classification loss, perform maximum entropy training on the classifier and minimum entropy training on the feature extractor. This process is continuously iterated to ensure the performance of the model.

6. The method according to claim 1, characterized in that Applied to the blast furnace ironmaking production process with multiple operating conditions switching, variable feed types, unstable ore quality, large fluctuations in data distribution, and few data labels.

Images