A blast furnace fault self-healing control method based on a generative adversarial framework

Abstract

The application discloses a blast furnace fault self-healing control method based on a generative adversarial framework, which comprises data set label attribute grading, construction of a latent space conversion self-healing control model, model multi-path training, and control variable adjustment of blast furnace fault samples. First, data samples need to be collected, control variables are selected as sample labels, and then label attributes are graded to obtain self-made data sets. Then, the latent space conversion self-healing control model is constructed, the self-made data sets are input into the model, and the model is trained according to a target function constructed according to a multi-training path. Finally, the blast furnace fault samples are input into the trained model, the model outputs a scheme matrix showing the working conditions of the conversion samples under different operations, and only the operation corresponding to the healthy working condition needs to be selected. The application establishes a large blast furnace characteristic generalized model of the relationship between the control variables, sample fault characteristics and abnormal furnace conditions, and provides certain operation guidance for coping with abnormal furnace conditions.

Description

TECHNICAL FIELD

[0001] The present application relates to a self-healing control method based on a generative adversarial framework for a blast furnace ironmaking process. BACKGROUND

[0002] Steel is a kind of metal material widely used in the construction, machinery and transportation industries, which has high strength, good mechanical properties, rich resources, low cost and is suitable for large-scale production, so it is closely related to industrialization process. The blast furnace is the core key unit of the conversion of iron material flow in the steel production process. However, the blast furnace is prone to various abnormal situations under the long-term operation of high temperature and high pressure, which not only affects the normal operation of the blast furnace, but also causes economic losses. The traditional abnormal treatment method mainly relies on manual experience to adjust the air supply, coal ratio and other parameters to restore the normal furnace condition. However, this method lacks theoretical guidance, is low in efficiency, and the adjustment range is difficult to control. Therefore, the self-healing control of abnormal conditions in the blast furnace ironmaking process is a problem to be solved. The main meaning of self-healing control in industrial process is to adjust the input and output, operation instruction, etc. of the control loop through active control means when the industrial process deviates from the safe operation of abnormal conditions, so as to make the system away from the fault condition, recover from the abnormal condition to the normal condition, and improve the safety of the system.

[0003] With the continuous innovation of technology, there have been a large number of effective results for fault diagnosis of blast furnace ironmaking process, but the self-healing control method for blast furnace fault is still lacking of research. The technical problems and difficulties faced by the research of large blast furnace system mainly concentrate on the complexity and dynamics of large blast furnace system. First of all, the mechanism of blast furnace system is extremely complex, the operation condition is variable, and the data sample is unbalanced, which brings great challenges to self-healing control. Secondly, the process variables of each subsystem in the blast furnace system are coupled with each other, and adjusting the key operation parameters such as air temperature, air volume, etc. will affect other process parameters, and then affect the whole furnace condition. Moreover, the large blast furnace is a complex dynamic system with insufficient knowledge and insufficient regulation means. The so-called "insufficient knowledge" mainly shows that there is a lack of precise and measurable control rules, a lack of complete and accurate mathematical model of blast furnace, and the internal state of the furnace cannot be directly measured. The "insufficient regulation means" is reflected in that only limited inputs such as air volume, air pressure and air temperature can be used as control variables to realize the control of the temperature distribution of each region in the furnace and the distribution of gas flow. Taking the pipeline travel fault of blast furnace ironmaking process as an example, the pipeline travel is one of the common abnormal furnace conditions of blast furnace ironmaking process, which is the manifestation of excessive development of air flow in a local cross section of blast furnace. Its formation and development are mainly due to the deterioration of raw fuel strength, the increase of powder, and the inadaptability of air volume and column permeability. The adjustment methods of pipeline travel are various, including reducing air, reducing oxygen, reducing coal, etc., but the implementation of these methods needs specific adjustment guidance, which puts high requirements on self-healing control technology.

[0004] Most iron smelting plants use control strategies based on human experience. The strategy mainly relies on human experience to judge the state of the blast furnace hearth, and takes corresponding control measures. This strategy has strong pertinence, but lacks theoretical support and is difficult to apply. The existing blast furnace self-healing control method constructs a self-healing control module based on a dynamic Bayesian network, obtains the type and severity of the fault through discretization threshold and dynamic reasoning, and then reasons the output of the self-healing control strategy and the adjustment value of the control variable, to realize fault diagnosis and automatic control. However, this method requires a large amount of expert knowledge. Modern blast furnace systems are equipped with a large number of sensors to monitor and collect various parameters and variables in the blast furnace production process. A large amount of data generated by these sensors can be converted into meaningful information through feature extraction methods to help the blast furnace system achieve self-healing control. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the present application provides a blast furnace fault self-healing control method based on a generative adversarial framework, comprising the following steps: data set label attribute grading, constructing a latent space conversion self-healing control model, model multi-path training, and control variable adjustment of blast furnace fault samples.

[0006] The present application is dedicated to the self-healing control of blast furnace iron smelting process faults, and proposes a blast furnace fault self-healing control method based on a generative adversarial framework to provide certain operation guidance for abnormal furnace conditions. First, data samples need to be collected, control variables are selected as sample labels, and then label attributes are graded to obtain self-made data sets. Then, a latent space conversion self-healing control model is constructed, the self-made data set is input into the model, and the model is trained according to the target function of multiple training paths. Finally, the blast furnace fault sample is input into the trained model, and the model outputs a scheme matrix showing the working conditions of the converted sample under different operations, and only the healthy working condition corresponding operation needs to be selected.

[0007] A blast furnace fault self-healing control method based on a generative adversarial framework, comprising the following steps: data set label attribute grading, constructing a latent space conversion self-healing control model, model multi-path training, and control variable adjustment of blast furnace fault samples.

[0008] The data set label attribute grading comprises the following steps:

[0009] (1.1) Collect data samples: according to the worker's log, collect blast furnace samples under normal furnace conditions and blast furnace samples when a fault occurs;

[0010] (1.2) Select control variables as sample labels: There are many process variables in the blast furnace ironmaking process. Based on the operation manual for handling blast furnace abnormalities, select control variables that regulate the blast furnace ironmaking process from the process variables, such as cold blast flow, hot blast temperature, and set coal injection amount. At the same time, select the control variables that need to be adjusted as the labels of the samples;

[0011] (1.3) Label attribute grading: After selecting the control variables as sample labels, the samples need to be divided into different data sets according to their measured values, that is, attribute grading. Because the cold blast flow and hot blast temperature are continuous values, they are graded into 0, 1, 2, and several other attributes, so that the label attributes are exclusive. The grading includes the following steps:

[0012] (1.3.1) Obtain the maximum and minimum values of the measured values of the control variables in the training set;

[0013] (1.3.2) After equal division, different grades are assigned. To obtain more detailed blast furnace self-healing control adjustment guidance, more attribute grades are divided;

[0014] (1.4) Make data sets: The graded cold blast flow, hot blast temperature, and set coal injection amount are used as the labels of the samples, and each sub-data set is divided to obtain a sample list file of different attributes.

[0015] The construction of the latent space conversion self-healing control model includes the following steps:

[0016] The proposed latent space conversion self-healing control model has six neural network modules, namely, encoder, generator, discriminator, extractor, converter, and mapper. On the basis of the generative adversarial network, three modules are added to control specific features. The extractor and the mapper are used to extract the feature code of the specific feature, and then the converter is used to modify the feature in the latent space;

[0017] (2.1) The function of the encoder is to encode the input sample into a latent space feature representation;

[0018] (2.2) The function of the generator is to decode the latent space feature representation into a blast furnace sample;

[0019] (2.3) The function of the discriminator is to judge the authenticity of the blast furnace sample and the working condition;

[0020] (2.4) The extractor is used to extract the feature code related to the label from the blast furnace sample. Specifically, the input of the extractor is x i，j , and the output is the extracted feature code. Here, i represents the number of labels, and j is the attribute number of a certain label. Each label has a group of attributes, but the attributes are only a rough discrete description of the label, so the feature code s i，jThe feature code is used to represent the specific manifestation of a specific label in the sample. Unlike the attribute, which simply describes the label, the feature code can truly reflect the subtle changes of the label in the sample. Because the attribute is only discrete annotation, while the feature code is continuous, it can better represent the subtle features of the label in the sample.

[0021] (2.5) The function of the converter is to control the specific label information of the feature based on the feature code, thereby modifying the feature. Specifically, it is assumed that the encoder extracts a feature vector of a blast furnace sample, which contains different label information. These label information is encoded in different dimensions of the feature vector. Then the feature vector and a specified feature code are input into the converter module. The converter contains multiple sets of adaptive instance normalization layers, corresponding to different labels. The label information of the input feature code will trigger the activation of the corresponding adaptive instance normalization layer, so as to only manipulate the part of the feature vector related to the label.

[0022] (2.6) The function of the mapper is to generate feature codes from random noise. Specifically, it has multiple sets of fully connected layers. First, according to the label i index, select the specific fully connected layer in front of the first layer fully connected layer, each set of fully connected layer corresponds to a label i, for the input label i, select the fully connected layer corresponding to it, and do not trigger other irrelevant fully connected layers. This can ensure that the feature code corresponds to a unique label. Then in the middle layer of the label corresponding fully connected layer, according to the attribute j index again, so that a certain layer performs specific conversion for a certain attribute. This makes the generated feature code contain the information of the specific attribute under the label.

[0023] The multi-path training of the model includes the following steps:

[0024] The collected data is used to train the model, so that it can learn the mapping relationship between the input fault sample and the control variable. The model needs to be trained according to three paths.

[0025] (3.1) The role of the non-conversion path is to reconstruct the input sample, ensuring the consistency between the encoder and the generator. The non-conversion path includes the following steps:

[0026] (3.1.1) input blast furnace sample x i，j , where i is the label and j is the attribute corresponding to the label;

[0027] (3.1.2) encode x i，j into feature e through the encoder E;

[0028] (3.1.3) input the feature e into the generator G to reconstruct the reconstructed sample x' i，j ;

[0029] (3.1.4) calculate xi，j and x′ i，j The reconstruction loss between them.

[0030] (3.2) The self-transformation path uses feature codes extracted from the same input sample for transformation, ensuring that the feature codes contain details of the input sample. The self-transformation path includes the following steps:

[0031] (3.2.1) Input the same blast furnace sample x i，j ;

[0032] (3.2.2) Use extractor F to extract from x i，j Extract the feature code s corresponding to label i. i，j ;

[0033] (3.2.3) Combine the sample features e and the extracted feature codes s i，j Input to converter T together;

[0034] (3.2.4) The transformed features are used by generator G to obtain the reconstructed sample x″. i，j ;

[0035] (3.2.5) Calculate x″ i，j With x i，j The reconstruction loss between them.

[0036] (3.3) The cyclic transformation path uses the generated feature code and the extracted feature code for a cyclically consistent transformation, ensuring that they both contain accurate label details. This guarantees that the mapper can learn the extractor and generate the required feature code. The cyclic transformation path includes the following steps:

[0037] (3.3.1) Randomly sample a target attribute

[0038] (3.3.2) Using mapper M according to Generate the corresponding target feature code

[0039] (3.3.3) x i，j Feature e and generated feature code The input transformer T, the transformed features are then processed by G to generate feature transformation samples after a self-healing operation.

[0040] (3.3.4) will Features and extracted raw feature codes s i，j The input converter performs an inverse conversion to obtain the cyclic reconstruction sample x″′. i，j ;

[0041] (3.3.5) Calculation The adversarial loss of x i，j The reconstruction loss between x i，j and x i，j and the feature code extraction consistency loss between the generated feature code and the extracted feature code .

[0042] (3.4) Then, the reconstruction loss, the adversarial loss and the feature code extraction consistency loss calculated by the three paths are all added together, which is the objective function of the model.

[0043] The control variable adjustment of the blast furnace fault sample comprises the following steps:

[0044] (4.1) The blast furnace fault sample is input into the encoder E to obtain the fault feature e;

[0045] (4.2) The fault feature e is converted by multiple converters, that is, the attributes of multiple labels are modified at the same time. According to the different levels of different labels, multiple operation schemes are obtained for converting the fault feature;

[0046] (4.3) The converted feature is decoded into a feature conversion sample by G. Multiple operation schemes can output a scheme matrix by the model to show the feature conversion sample working conditions under different operation schemes. If a feature conversion sample is a healthy sample, the corresponding operation scheme is a self-healing control operation.

[0047] The beneficial effects of the present application are:

[0048] The method of the present application overcomes the defects of traditional manual experience type adjustment of blast furnace faults, and proposes a data-driven self-healing control method for blast furnace faults based on a generative adversarial framework. The method establishes a mapping relationship model between the blast furnace running state and the control variable, can automatically provide a feasible adjustment scheme according to the fault condition, and improves the efficiency and accuracy of the self-healing control. BRIEF DESCRIPTION OF DRAWINGS

[0049] Figure 1 It is a self-healing control model training path diagram for the blast furnace ironmaking process of the present application.

[0050] Figure 2 It is a sample feature conversion diagram for the self-healing control model of the blast furnace ironmaking process of the present application.

[0051] Figure 3 It is a case 1 diagram of the operation scheme matrix output by the self-healing control model of the present application.

[0052] Figure 4 It is a case 2 diagram of the operation scheme matrix output by the self-healing control model of the present application. DETAILED DESCRIPTION

[0053] The application is further described below in connection with the accompanying drawings and examples.

[0054] Taking the blast furnace pipeline stroke failure as an example, a blast furnace failure self-recovery control method based on a generative adversarial framework is introduced, including the following steps:

[0055] (1) Collecting data samples: According to the worker's log, 5000 blast furnace samples under normal furnace conditions and 5000 blast furnace samples when the pipeline stroke failure occurs are collected.

[0056] (2) Selecting control variables as sample labels: The blast furnace ironmaking system can collect many process variables. Taking the pipeline stroke failure as an example, based on the operation manual for handling blast furnace abnormalities, it is necessary to reduce the air, reduce the coal, and reduce the temperature, so the air volume, the air temperature, and the coal injection amount are important parameters for adjusting the blast furnace sample. Therefore, the cold air flow, the hot air temperature, and the set coal injection amount are used to represent the air volume, the air temperature, and the coal injection amount. At the same time, the selected control variables that need to be adjusted are also used as the labels of the samples.

[0057] (3) Label attribute grading: After selecting the three control variables as sample labels, the samples need to be divided into different data sets according to their measurement values, that is, attribute grading. Because the cold air flow and the hot air temperature are continuous values, they are graded into 0, 1, 2, and several other attributes, so that the label attributes have exclusivity. It should be noted that at this time, the blast furnace sample only retains other process variables, and the three control variables as labels will not appear in the sample. The specific measure of grading is to obtain the maximum and minimum values of the measurement values of the three control variables in the training set, and then grade them, such as the measurement value of the cold air flow being between 44-50, then setting 44-46 as level 0, 46-48 as level 1, and 48-50 as level 2. To obtain more detailed blast furnace self-recovery control adjustment guidance, more attribute levels need to be divided.

[0058] (4) Making data sets: Write code to use the graded cold air flow, hot air temperature, and set coal injection amount as sample labels, and divide each sub-data set, such as each variable having three attributes, then a total of 9 sub-data sets are divided. And get the sample list file of different attributes.

[0059] (5) Constructing the latent space conversion self-healing control model: the proposed latent space conversion self-healing control model has six neural network modules, which are encoder, generator, discriminator, extractor, converter and mapper. On the basis of the generative adversarial network, three modules are added to control specific features. The extractor and the mapper are used to extract the feature code of the specific feature, and then the converter is used to modify the feature in the latent space. The function of the encoder is to encode the input sample into a feature representation. The function of the generator is to decode the feature representation into a blast furnace sample. The function of the discriminator is to judge the authenticity of the blast furnace sample and the working condition.

[0060] The extractor is used to extract the label-related feature code from the sample. Specifically, its input is the blast furnace sample x i，j , and the output is the extracted feature code. Here, i represents the number of labels, and j is the attribute number of a certain label. Each label has a set of attributes, such as level 0 and level 1 attributes of cold wind flow. It should be noted that attribute j cannot represent the specific performance of label i in the sample. The attribute is only a rough discrete description of the label, and level 1 cannot represent the specific value of the cold wind flow. Therefore, a feature code s i，j is introduced to represent the specific performance of a specific label in the blast furnace sample. Because the attribute is only a discrete label, and the feature code is continuous, unlike the simple description of the attribute to the label, the feature code can truly reflect the subtle changes of the label in the sample.

[0061] The function of the converter is to control the specific label information of the feature based on the feature code, so as to modify the feature. Specifically, suppose the encoder extracts a feature vector of a blast furnace sample, which contains the label information of the sample, such as cold wind flow, hot wind temperature, etc. These label information is encoded in different dimensions of the feature vector. Then input the feature vector and a specified feature code into the converter module. The converter has a learnable adaptive instance normalization layer that injects information from the feature code into the corresponding dimension of the feature vector, thereby changing the features related only to the "cold wind flow" label. The converter contains multiple adaptive instance normalization layers, corresponding to different labels. The label information of the input feature code will trigger the activation of the corresponding adaptive instance normalization layer, thereby only manipulating the part of the feature vector related to the label.

[0062] The function of the mapper is to generate the feature code from the random noise. Specifically, it has multiple sets of fully connected layers. First, a specific fully connected layer is selected according to the label i index before the first layer of fully connected layers, and each set corresponds to a label i, such as the first set controls the cold air flow, and the second set controls the hot air temperature. For the input label i, the corresponding set of fully connected layers is selected, and other irrelevant fully connected layers are not triggered. In this way, it can be ensured that the feature code corresponds to a unique label. Then, in the middle layer of the fully connected layer corresponding to the label, according to the attribute j index again, so that a certain layer performs a specific conversion for a certain attribute. This makes the generated feature code contain the information of the specific attribute under the label.

[0063] (6) After the self-healing control model is constructed, it needs to be trained with the blast furnace samples, so that it can learn the mapping relationship between the input fault samples and the control variables. The model needs to be trained according to the three paths in Figure 1 .

[0064] (6.1) The function of the non-conversion path is to reconstruct the input sample and ensure the consistency between the encoder and the generator. The non-conversion path includes the following steps:

[0065] (6.1.1) Input the blast furnace sample x i，j , where i is the label and j is the attribute corresponding to the label;

[0066] (6.1.2) Encode x i，j into feature e through the encoder E;

[0067] (6.1.3) Input the feature e into the generator G to reconstruct the reconstructed sample x' i，j ;

[0068] (6.1.4) Calculate the reconstruction loss L i，j between x i，j and x' rec1 = E i，j，x [||x′ i，j -x i，j ||1].

[0069] (6.2) The self-conversion path uses the feature code extracted from the same input sample to convert, ensuring that the feature code contains the details of the input sample. The self-conversion path includes the following steps:

[0070] (6.2.1) Input the same blast furnace sample x i，j ;

[0071] (6.2.2) Extract the feature code s i，j corresponding to the label i from x i，j using the extractor F;

[0072] (6.2.3) input the features e of the sample and the extracted feature code s into the converter T; i，j a converter T;

[0073] (6.2.4) the converted features are reconstructed into a sample x" by the generator G i，j ;

[0074] (6.2.5) calculate the reconstruction loss L between x" i，j and x i，j ; rec2 = E i，j，x [||x" i，j -x i，j ||1].

[0075] (6.3) the cycle conversion path uses the generated feature code and the extracted feature code to perform cycle-consistency conversion, ensuring that they both contain accurate label details. This guarantees that the mapper can learn the extractor to generate the required feature code. The cycle conversion path includes the following steps:

[0076] (6.3.1) randomly sample a target attribute

[0077] (6.3.2) use the mapper M to generate the corresponding target feature code according to

[0078] (6.3.3) input the features e of x i，j and the generated feature code into the converter T, and the converted features are generated into the feature conversion sample after self-healing operation by G

[0079] (6.3.4) input the features of and the extracted original feature code s i，j into the converter for inverse conversion, to obtain the cycle reconstruction sample x" i，j ;

[0080] (6.3.5) calculate the adversarial loss between x i，j and x" i，j , the reconstruction loss between x" i，j and x i，j , and the feature consistency loss between the extracted feature code and the generated feature code ;

[0081] Then the reconstruction loss, the adversarial loss and the feature consistency loss of the three paths are all added together, which is the objective function of the model λ rec and λ feat is a hyperparameter that controls the relative importance of the corresponding loss function compared to the adversarial loss.

[0082] (7) After the model training is completed, the control variable adjustment scheme of the blast furnace fault input sample can be obtained. The sample conversion path of the model self-healing stage is shown in Figure 2 , the fault feature e is obtained after the certain blast furnace pipeline stroke fault sample is input into the encoder E. The different attributes of each label are operated, and the fault feature e is converted by multiple converters, that is, the attributes of multiple labels are modified at the same time. According to the different levels of different labels, multiple operation schemes are obtained for converting the fault feature. The converted feature is decoded into a feature conversion sample by G, and the working condition of the feature conversion sample is given by the discriminator. Multiple operation schemes can output a scheme matrix by the model to show the working condition of the feature conversion sample under different operation schemes. If a certain feature conversion sample is a healthy sample, the corresponding operation scheme is a self-healing control operation. The priority of the control variable is determined based on expert knowledge, which is to reduce the air flow, reduce the coal flow and reduce the temperature. As shown in Figure 3 , the cold air flow is adjusted first, and when the cold air flow is reduced to level 0, the sample is judged to be healthy, and only the air flow can be operated. As shown in Figure 4 , even if the cold air flow is adjusted to level 0, the sample is still a fault, and the coal injection amount needs to be continuously modified. When the coal injection amount is reduced to level 1, the sample can be judged to be healthy, and the operation scheme needs to modify two control variables.

[0083] The technical features of the above-described embodiments can be further combined. In order to make the description simple, not all possible combinations of the technical features in the above-described embodiments are described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the description.

[0084] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it cannot be understood as a limitation on the scope of the application. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. The protection scope of the present application is given by the appended claims and any equivalent technical solutions thereof.

Claims

1. A blast furnace fault self-healing control method based on a generative adversarial framework, characterized in that, The method comprises the following steps: The data set label attribute grading, the construction latent space conversion self-healing control model, the model multi-path training, the control variable adjustment of blast furnace fault sample; First, the data sample needs to be collected, the control variable is selected as the sample label, and then the label attribute is graded to obtain the self-made data set; then the latent space conversion self-healing control model is constructed, the self-made data set is input into the model, and the model is trained according to the target function of the multi-training path; Finally, the blast furnace fault sample is input into the trained model, the model outputs a scheme matrix to display the working conditions of the conversion sample under different operations, and the operation corresponding to the healthy working condition is selected; The model multi-path training comprises the following steps: The collected data is used to train the model, so that the model can learn the mapping relationship between the input fault sample and the control variable; the model is trained according to three paths; (3.1) The non-conversion path comprises the following steps: (3.1.1) input blast samples x i，j where i is a label and j is the attribute corresponding to the label. (3.1.2) x i，j encoded as a feature e by the encoder E; (3.1.3) inputting the feature e into the generator G to reconstruct to obtain a reconstructed sample x' i，j ; (3.1.4) Compute x i，j and the reconstruction loss between x′ i，j and x. (3.2) The self-conversion path uses the feature code extracted from the same input sample to perform conversion, so as to ensure that the feature code contains the details of the input sample; the self-conversion path comprises the following steps: (3.2.1) inputting the same blast furnace sample x i，j ; (3.2.2) using the extractor F to extract the feature code s corresponding to the tag i from x i，j i，j ;​ (3.2.3) the feature e of the sample and the extracted feature code s are compared with each other i，j The input converter T; (3.2.4) The transformed features are used by the generator G to reconstruct the sample x" i，j ; (3.2.5) Compute x" i，j between x i，j and the reconstruction loss; (3.3) The cyclic conversion path uses the generated feature code and the extracted feature code to perform cyclic consistency conversion, so as to ensure that they both contain accurate label details; the cyclic conversion path comprises the following steps: (3.3.1) Randomly sampling a target attribute (3.3.2) using the mapper M according to generating a corresponding target feature code (3.3.3) x i，j characteristics e and generated characteristic codes input converter T, converted characteristics generate characteristics after self-recovery operation of the feature conversion sample (3.3.4) the features of the characteristic and the extracted original feature code s i，j are input into the converter for inverse conversion to obtain the cyclically reconstructed samples x′′′ i，j ; (3.3.5) Computing the adversarial loss between x i，j and x i，j , the reconstruction loss between x i，j and x , and the feature consistency loss between the extracted feature code and the generated feature code (3.4) Then, the reconstruction loss, the adversarial loss and the feature consistency loss calculated by the three paths are all added, that is, the target function of the model.

2. The method of claim 1, wherein, The data set label attribute grading comprises the following steps: (1.1) Collecting data samples: according to the worker's log, collecting blast furnace samples under normal furnace conditions and blast furnace samples when faults occur; (1.2) Selecting the control variable as the sample label: selecting the control variable for adjusting the blast furnace ironmaking process from the process variable, including the cold blast flow, the hot blast temperature and the set coal injection amount; at the same time, the control variable selected as needing to be adjusted is also used as the label of the sample; (1.3) Label attribute grading: after selecting the control variable as the sample label, the sample is divided into different data sets according to the measurement value, that is, the attribute is graded; according to the continuous values of the cold blast flow and the hot blast temperature, the attribute is graded into several attributes, so that the label attribute has exclusivity; the steps of grading are as follows: (1.3.1) Obtain the maximum and minimum values of the measurement values of the control variables in the training set; (1.3.2) After equal division, different levels are assigned; (1.4) Making data set: taking the graded cold blast flow, hot blast temperature and set coal injection amount as the label of the sample, dividing into each sub-data set, and obtaining the sample list file of different attributes.

3. The method of claim 1, wherein, The construction latent space conversion self-healing control model comprises the following steps: The proposed latent space conversion self-healing control model has six neural network modules, which are respectively an encoder, a generator, a discriminator, an extractor, a converter and a mapper; on the basis of the generative adversarial network, three modules are added to control specific features, the extractor and the mapper are used to extract the feature code of the specific feature, and then the converter is used to modify the feature in the latent space; (2.1) The function of the encoder is to encode the input sample into a latent space feature representation; (2.2) The function of the generator is to decode the latent space feature representation into a blast furnace sample; (2.3) The function of the discriminator is to judge the authenticity of the blast furnace sample and the working condition; (2.4) The extractor is used to extract label-related feature codes from the blast furnace sample; Input is x i，j , output is extracted feature code, here i represents the number of labels, j is the attribute number of a certain label; each label has a set of attributes, and the feature code s i，j is used to represent the specific performance related to a specific label in the sample; (2.5) The function of the converter is to control the specific label information of the feature based on the feature code, thereby modifying the feature; the encoder extracts a feature vector of a blast furnace sample, which contains different label information; these label information is encoded in different dimensions of the feature vector; then the feature vector and a specified feature code are input into the converter module; the converter contains multiple sets of adaptive instance normalization layers, each corresponding to a different label; The label information of the input feature code will trigger the activation of the corresponding adaptive instance normalization layer, so as to only manipulate the part of the feature vector related to the label; (2.6) The function of the mapper is to generate feature codes from random noise; There are multiple fully connected layers, first according to the label i index to select the specific fully connected layer in front of the first layer fully connected layer, each set of fully connected layers corresponds to a label i, for the input label i, select the corresponding set of fully connected layers, without triggering other irrelevant fully connected layers; then in the middle layer of the label corresponding fully connected layer, according to the attribute j index again, so that a certain layer performs specific conversion for a certain attribute; so that the generated feature code contains the information of the specific attribute under the label.

4. The method of claim 1, wherein, The control variable adjustment of the blast furnace fault sample includes the following steps: (4.1) input the blast furnace fault sample into the encoder E to obtain the fault feature e; (4.2) convert the fault feature e with multiple converters, that is, modify the attributes of multiple labels at the same time; according to the different levels of attributes of different labels, obtain multiple operation schemes for converting the fault feature; (4.3) decode the converted feature through G into a feature conversion sample; multiple operation schemes output a scheme matrix through the model, showing the feature conversion sample working condition under different operation schemes; if a feature conversion sample is a healthy sample, its corresponding operation scheme is a self-healing control operation.

Images