CN113219942B - A blast furnace fault diagnosis method based on weighted joint distributed adaptive neural network - Google Patents

Abstract

The invention discloses a blast furnace fault diagnosis method based on a weighted joint distribution adaptive neural network. Firstly, performing first-time feature extraction on the historical data and the data to be detected of the blast furnace by using a deep neural network and generating a label of the data to be detected. Based on the label value, the ratio of the to-be-measured data of the blast furnace to the number of samples of the corresponding fault category in the historical data is calculated, and the obtained ratio is used as the corresponding weight of each type of fault of the blast furnace to be combined with a joint distribution adaptation method. And finishing the extraction of the second characteristic by weighted joint distribution adaptation and obtaining a new label value. And finally, carrying out iterative solution on the process of generating labels in the weighted joint distribution adaptation, calculating weights and updating parameters to obtain a fault diagnosis result. The invention not only utilizes the deep neural network to improve the diagnosis precision, but also solves the problem of low accuracy of the traditional fault diagnosis method caused by less fault samples of the blast furnace and larger fluctuation of data distribution along with the change of working conditions through the adaptation of combined distribution and prior distribution.

Description

A blast furnace fault diagnosis method based on weighted joint distributed adaptive neural network

technical field

The invention belongs to the field of industrial process monitoring, modeling and simulation, in particular to a blast furnace fault diagnosis method based on weighted joint distribution adaptive neural network.

Background technique

Blast furnace ironmaking is the core unit of ferrite flow conversion, and it is the link with the largest energy consumption and the highest production cost in the iron and steel manufacturing process. With the continuous progress of technology and technology in the blast furnace ironmaking process, and the continuous development of instrumentation and automation technology, modern blast furnace ironmaking has the characteristics of large scale, complex structure, strong coupling between production units and huge investment. Abnormal fluctuations (or accidents) in the blast furnace ironmaking process are not discovered in time, which often leads to a serious decline in product quality, or delays the normal execution of production plans, resulting in huge economic losses and even casualties. Therefore, fault diagnosis of blast furnace is of great significance to ensure the safe and efficient production of blast furnace.

In the actual production process of blast furnace ironmaking, in order to avoid serious consequences, the operator will adjust the air supply system, distribution system or furnace heat system to avoid failures when there are certain signs of failure in the operation of the blast furnace system. Therefore, under the existing operating system and operating conditions, the fault diagnosis system for blast furnace construction is faced with problems such as few fault samples, unbalanced data, missing labels, and expensive and time-consuming labeling samples. In addition, because the origin of raw materials is not fixed, most domestic steel mills use the form of "Baijia Mine" for blast furnace ironmaking. At different times, the type of feed and its ratio will change significantly; secondly, there are various switching conditions during the production and operation of the blast furnace. These factors all cause the blast furnace data to change with time, the data distribution fluctuates greatly, and there is a distribution difference between the training data and the data to be tested, which affects the reliability and accuracy of fault diagnosis.

At present, the fault diagnosis methods used in blast furnaces can be roughly divided into two types, namely expert systems and data-driven intelligent fault diagnosis methods. Expert systems have higher requirements for prior knowledge such as relevant knowledge and rules, while blast furnaces involve physical It is extremely complex with chemical reactions, and it is difficult to know the exact situation of the actual internal reaction. And with the widespread use of various intelligent instruments and control equipment in the distributed control system in modern industrial processes, a large amount of process data is collected and stored. However, these data containing process operating state information are often not used effectively in expert systems.

On the other hand, there are two prerequisites for the successful application of traditional data-driven intelligent fault diagnosis methods: 1) a large amount of labeled data, and 2) the training and testing data come from the same data distribution. However, in the blast furnace production process, there are very few fault samples with labels, and it is difficult to obtain, and due to the different grades of ore raw materials and production conditions, the data will fluctuate greatly, resulting in the training data and test data often cannot meet the requirements. conditions for the same distribution. Therefore, the existing abnormal furnace condition diagnosis methods are still far from ideal practical applications.

SUMMARY OF THE INVENTION

In order to overcome the deficiencies of the prior art, the purpose of the present invention is to provide a blast furnace fault diagnosis method based on a weighted joint distribution adaptive neural network. The method firstly uses a neural network to perform the first feature extraction on the historical data of the blast furnace and the data to be measured, and generates the initial label of the data to be measured. Based on this tag value, the ratio of the number of samples of the blast furnace to be tested and the corresponding fault category in the historical data is calculated, and the obtained ratio is used as the corresponding weight of various types of blast furnace faults and combined with the joint distribution adaptation method. The second feature extraction is completed through weighted joint distribution adaptation and a new label value is obtained. Finally, the process of generating labels, calculating weights and updating parameters in weighted joint distribution adaptation is iteratively solved to obtain fault diagnosis results. It can be widely used in industrial systems that require high reliability and accuracy for fault diagnosis.

A blast furnace fault diagnosis method based on weighted joint distribution adaptive neural network, the steps are as follows:

Step 1: Use the blast furnace historical data and blast furnace test data to train the weights of the deep neural network. The data value obtained by the last fully connected layer in the neural network is the extracted feature. The fault diagnosis error of the blast furnace historical data The sum of the distances between the extracted features of the two sets of data is used as the loss function, and the weight is fixed after the training reaches a preset number of iterations or the loss function is lower than the preset value;

Step 2: Use the fixed neural network in Step 1 to perform the first feature extraction on the blast furnace historical data and the data to be measured, and generate an initial label value for the blast furnace to be measured data, that is, to perform feature migration between the historical data of the blast furnace and the data to be measured. On the basis of using the blast furnace fault diagnosis knowledge learned from the blast furnace historical data and the data to be measured, a nonlinear mapping from blast furnace process variables to blast furnace fault categories is initially formed;

Step 3: Based on the label value of the data to be tested, calculate the proportion of each fault category in the data to be tested and the historical data of the blast furnace respectively, and compare the ratio of the data to be tested of the blast furnace with the proportion of the corresponding category in the historical data to obtain the ratio. The prior distribution weight of the category of the blast furnace data to be tested is multiplied by the corresponding blast furnace historical data features extracted in step 2, and together with the characteristics of the blast furnace to be tested data extracted in step 2 to form a feature variable matrix, after weighting, The fault category distribution of the blast furnace historical data tends to be consistent with the fault category distribution of the blast furnace to-be-measured data, so as to realize the priori category distribution adaptation of the two;

Step 4: Introduce the kernel method, map the feature variables, obtain new feature variables, and transform the feature variables in the kernel space, so that the feature vectors extracted from the historical data of the blast furnace and the data to be tested are distributed in the marginal distribution and conditions. The sum of the distances on the distribution is the smallest. Therefore, the joint distribution adaptation of the blast furnace historical data and the data to be measured is realized by the method of the kernel method and the transformation matrix;

Step 5: Finally, the transformed feature variable is used as the input of the classifier. The connection weight between the feature variable and the classifier needs to be trained with the classification accuracy rate as the objective function. After convergence, the classification result of the classifier to be tested is The blast furnace failure category is assigned as a new tag value to the data to be tested;

Step 6: Repeat steps 3 to 5 until the distance between the eigenvectors of the blast furnace historical data and the data to be tested on the joint distribution and the classification accuracy rate are both stable, then fix the model parameters, and perform the analysis on the data to be tested. Discriminate processing and generate fault diagnosis results.

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 an output layer, and the input layer is the input layer of blast furnace process variable parameters, including air permeability index, cold air flow, hot air flow, top Pressure, cold air pressure, hot air pressure and other industrial process parameters that characterize the production state of the blast furnace, and the output layer is the blast furnace fault category layer, including the blast furnace that is related to the blast furnace production process, such as difficult operation, suspended material, pipeline, collapsed material, furnace heat, and furnace cooling. Fault, the function of the hidden layer is to establish a nonlinear mapping from blast furnace process variables to blast furnace fault categories, so that blast furnace fault diagnosis knowledge can be learned from blast furnace historical fault data, and a blast furnace fault diagnosis model can be established. The neurons in the same layer are not connected, the neurons between layers are fully connected, and each connection has a weight, which represents the strength of the connection between neurons. For different industrial application fields, the requirements for the number of hidden layers of deep neural networks 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 a deep neural network is:

为神经网络第i层第j个隐藏层单元的输出，记h_i为神经网络第i层，则h₀为神经网络输入层，h_k+1为神经网络输出层；j的取值根据网络第i层的神经元的个数决定，记第i层的神经元个数为z_i，则每层j的取值为1到z_i；W(i,j)为第i层第j个神经元对应的权值矩阵；

为第i层第j个神经元对应的偏置项，b_k+1为输出层单元对应的偏置项；y代表神经网络的输出,M为高炉历史数据的样本总数，记N为高炉待测数据的样本总数，f(·)和g(·)分别是隐层单元和输出单元的激活函数，

代表第i个样本在输出层神经元中的最大值，s_j代表输出层中第j个神经元的数值，将全连接层的数据作为特征向量进行提取，将高炉历史数据与待测数据提取出特征向量分别记为x^s与x^t，将高炉历史数据的故障诊断误差与两组数据所提取特征之间距离的总和作为损失函数，即为下式：in,

is the output of the jth hidden layer unit of the i-th layer of the neural network, and denote h _i as the i-th layer of the neural network, then h ₀ is the input layer of the neural network, and h _k+1 is the output layer of the neural network; the value of j is based on the network The number of neurons in the i-th layer is determined, and the number of neurons in the i-th layer is denoted as _zi , then the value of each layer j is 1 to _zi ; W(i,j) is the j-th layer of the i-th layer The weight matrix corresponding to the neuron;

is the bias term corresponding to the jth neuron in the i-th layer, b _k+1 is the bias term corresponding to the output layer unit; y represents the output of the neural network, M is the total number of samples of the blast furnace historical data, and N is the blast furnace waiting time The total number of samples of the test data, f( ) and g( ) are the activation functions of the hidden layer unit and the output unit, respectively,

Represents the maximum value of the i-th sample in the neurons of the output layer, s _j represents the value of the j-th neuron in the output layer, extracts the data of the fully connected layer as a feature vector, and extracts the historical data of the blast furnace and the data to be tested. The feature vectors are recorded as x ^s and x ^t respectively, and the sum of the fault diagnosis error of the blast furnace historical data and the distance between the features extracted from the two sets of data is used as the loss function, which is the following formula:

The steps of weighting described in step 3 are as follows: the blast furnace faults are classified as C, and c represents the type of blast furnace fault. When c is a real number from 1 to C, it represents the corresponding specific fault type, such as: pipeline, down, difficult to travel, suspended. materials, etc., M is the total number of samples of blast furnace historical data, in which the number of samples belonging to type c fault type is M _C , correspondingly, N is the total number of samples of blast furnace data to be measured, of which the number of samples belonging to type c fault type is N _C , the label value of the historical data is recorded as y ^s , the label value of the data to be measured is recorded as y ^t , the distribution of the historical data of the blast furnace and the data to be measured are recorded as _ps ( ) and _pt ( ) respectively, the historical data of the blast furnace is recorded as ps ( ) and pt ( ) The proportions of various fault samples in the data to be tested are as follows:

Then the corresponding weights of various fault data in the blast furnace historical data are:

After multiplying the weight by the historical data of the blast furnace, the prior distribution of the historical data of the blast furnace is:

Therefore, the fault category distribution of the blast furnace historical data tends to be consistent with the category prior distribution of the blast furnace data to be tested, and the prior category distribution adaptation of the two is realized. After weighting, the characteristic matrix X ^s of the blast furnace historical data and the characteristic matrix X ^t of the data to be measured form the characteristic matrix X.

The steps of adapting the kernel mapping and joint distribution described in step 4 are as follows: selecting a kernel function such as a Gaussian kernel function, and performing nonlinear mapping on the features, namely:

为非线性映射函数，则高炉历史数据与待测数据在核空间内在联合分布上的距离之为：in

is a nonlinear mapping function, then the distance between the historical data of the blast furnace and the data to be tested in the joint distribution in the kernel space is:

Introduce the kernel matrix K:

in:

Assuming that the required transformation matrix is W, the joint distribution distance between the blast furnace historical data and the data to be measured is:

Combined with the kernel matrix, the minimization problem of the above formula can be transformed into:

st W ^T KHK ^T W=I

Where: H=I _M+N -1/(M+N)11 ^T

The solution of this equation, the transformation matrix W, can be obtained by eigenvalue decomposition.

The steps of iterative update described in step 6 are as follows: iteratively solve steps 3 to 5, that is, iteratively obtain a fault diagnosis result in the process of generating labels in weighted joint distribution adaptation, calculating weights and updating parameters.

Beneficial effects of the present invention:

Aiming at the characteristics and problems such as large data fluctuation, missing labels, and data imbalance in the blast furnace ironmaking process, a blast furnace fault diagnosis method based on weighted joint distribution adaptive neural network was constructed. It fully mines the knowledge contained in the data, and solves the problem that there is a lot of historical data in blast furnaces, but it is difficult to directly train the model for the data to be tested. level of intelligence.

Description of drawings

FIG. 1 shows a block flow diagram of the method of the present invention.

Figure 2 shows the original distribution of the data to be measured by t-sne visualization results.

FIG. 3 shows the visualization result of t-sne after the blast furnace fault classification is performed on the data to be measured by the method of the present invention.

Detailed ways

The present invention will be further described below with reference to the accompanying drawings.

The purpose of the present invention is to provide a blast furnace fault diagnosis method based on weighted joint distribution adaptive neural network. The flow chart is shown in Figure 1. Considering the nonlinear and non-Gaussian properties of blast furnace production information, the deep neural network can be used for infinite approximation. The first feature extraction is based on the advantages of the nonlinear function, and the marginal distribution distance between the blast furnace historical data and the data to be tested is measured by the maximum mean distance (MMD), and the sum of the classification accuracy and distance is used as the loss function. The training and weights are fixed. The fully connected layer of the deep neural network is used as the feature vector extracted by the neural network, and the classification result of the neural network is used as its initial label value. The ratio is weighted as a priori distribution weight of the category, and then the feature vector is mapped in the kernel space by introducing a kernel function, and the distance between the historical data of the blast furnace and the feature vector of the data to be tested in the joint distribution is weighted and combined with the maximum average distance. (WJMMD) is used to measure, and the distance is minimized by calculating the transformation matrix, and then a new classification result is obtained through the classifier (softmax). Finally, the process of generating labels, calculating weights and updating parameters in weighted joint distribution adaptation is iteratively solved to obtain fault diagnosis results. The method is helpful to realize knowledge and decision enhancement in blast furnace fault diagnosis, and ensure the reliability and accuracy of blast furnace fault diagnosis.

Next, the effectiveness of the method of the present invention is verified by using the blast furnace fault data collected by the No. 2 blast furnace of a steel plant. The blast furnace is divided into five parts: throat, shaft, waist, belly and hearth from top to bottom. Coke, ore and flux will undergo different changes in different parts of the furnace during the descending process until they reach the bottom of the hearth and are completely transformed. For molten iron and slag. Due to the large size of the blast furnace and the complex chemical reactions that take place in the furnace, it is extremely important to ensure its safe and smooth operation. Blast furnace failures are mainly divided into four categories: difficult, hanging, channeling, and collapsing. The data collected in the production process includes 29 parameters such as air permeability index, cold air flow, hot air flow, top pressure, cold air pressure, and hot air pressure. In actual production, the three-shift system is used to organize workers to monitor and operate the blast furnace ironmaking process, which consumes a lot of labor costs, and the control method is relatively extensive, mainly relying on several parameters to judge the furnace condition. It is difficult to diagnose the problems existing in the blast furnace operation process in time and carry out accurate control in time. The method of the invention can solve this problem to a certain extent and has practical application value.

Next, the implementation steps of the present invention will be described in detail in conjunction with the specific process.

1. Build a deep neural network for the first feature extraction and generate initial label values

(1) Construct a deep neural network by using the blast furnace historical data and unlabeled data to be tested. The structure is shown in the feature extraction part of the neural network in Figure 1. It includes three parts: the input layer, the hidden layer and the output layer. The input layer is the blast furnace process. Variable parameter input layer, including air permeability index, cold air flow, hot air flow, top pressure, cold air pressure, hot air pressure and other industrial process parameters that characterize blast furnace production status, and the output layer is the blast furnace fault category layer, including difficult to travel, suspended material, pipeline The function of the hidden layer is to establish a nonlinear mapping from blast furnace process variables to blast furnace fault categories, so that the blast furnace can be learned from the blast furnace historical fault data. Fault diagnosis knowledge, establish blast furnace fault diagnosis model. The neurons in the same layer are not connected, the neurons between layers are fully connected, and each connection has a weight, which represents the strength of the connection between neurons.

For different industrial application fields, the requirements for the number of hidden layers of deep neural networks 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 a deep neural network is:

为神经网络第i层第j个隐藏层单元的输出，记h_i为神经网络第i层，则h₀为神经网络输入层，h_k+1为神经网络输出层；j的取值根据网络第i层的神经元的个数决定，记第i层的神经元个数为z_i，则每层j的取值为1到z_i；W(i,j)为第i层第j个神经元对应的权值矩阵；

为第i层第j个神经元对应的偏置项，b_k+1为输出层单元对应的偏置项；y代表神经网络的输出,M为高炉历史数据的样本总数，记N为高炉待测数据的样本总数，f(·)和g(·)分别是隐层单元和输出单元的激活函数，

代表第i个样本在输出层神经元中的最大值，s_j代表输出层中第j个神经元的数值，将全连接层的数据作为特征向量进行提取，将高炉历史数据与待测数据提取出特征向量分别记为x^s与x^t，将高炉历史数据的故障诊断误差与两组数据所提取特征之间距离的总和作为损失函数，即为下式：in,

is the output of the jth hidden layer unit of the i-th layer of the neural network, and denote h _i as the i-th layer of the neural network, then h ₀ is the input layer of the neural network, and h _k+1 is the output layer of the neural network; the value of j is based on the network The number of neurons in the i-th layer is determined, and the number of neurons in the i-th layer is denoted as _zi , then the value of each layer j is 1 to _zi ; W(i,j) is the j-th layer of the i-th layer The weight matrix corresponding to the neuron;

is the bias term corresponding to the jth neuron in the i-th layer, b _k+1 is the bias term corresponding to the output layer unit; y represents the output of the neural network, M is the total number of samples of the blast furnace historical data, and N is the blast furnace waiting time The total number of samples of the test data, f( ) and g( ) are the activation functions of the hidden layer unit and the output unit, respectively,

Represents the maximum value of the i-th sample in the neurons of the output layer, s _j represents the value of the j-th neuron in the output layer, extracts the data of the fully connected layer as a feature vector, and extracts the historical data of the blast furnace and the data to be tested. The feature vectors are recorded as x ^s and x ^t respectively, and the sum of the fault diagnosis error of the blast furnace historical data and the distance between the features extracted from the two sets of data is used as the loss function, which is the following formula:

(2) Use the fixed neural network in (1) to perform the first feature extraction on the historical data and the data to be measured of the blast furnace and generate the initial label value for the data to be measured of the blast furnace, that is, the historical data of the blast furnace and the data to be measured are characterized. On the basis of migration, a nonlinear mapping from blast furnace process variables to blast furnace fault categories is initially formed by using blast furnace fault diagnosis knowledge learned from blast furnace historical data and data to be measured.

2. Weighting the extracted feature vector to complete the category prior distribution adaptation

The steps of weighting the eigenvectors extracted from the blast furnace historical data are as follows: record the blast furnace fault as a total of C types, and c represents the blast furnace fault type. When c is a real number from 1 to C, it represents the corresponding specific fault type, such as: pipeline, Downward, difficult, suspended material, etc., M is the total number of samples of blast furnace historical data, and the number of samples belonging to the type c fault type is M _C , correspondingly, N is the total number of samples of the blast furnace data to be measured, which belongs to the type c fault. The number of samples of the type is N _C , the label value of the historical data is recorded as y ^s , the label value of the data to be measured is recorded as y ^t , and the distribution of the historical data and the data to be measured are recorded as p _s ( ) and p _t ( ), the proportions of various fault samples in the blast furnace historical data and the data to be tested are:

Then the corresponding weights of various fault data in the blast furnace historical data are:

After multiplying the weight by the historical data of the blast furnace, the prior distribution of the historical data of the blast furnace is:

Therefore, the fault category distribution of the blast furnace historical data tends to be consistent with the category prior distribution of the blast furnace data to be tested, and the prior category distribution adaptation of the two is realized. After weighting, the characteristic matrix X ^s of the blast furnace historical data and the characteristic matrix X ^t of the data to be measured form the characteristic matrix X.

3. Kernel mapping and joint distribution adaptation of feature vectors

(1) Select a kernel function such as a Gaussian kernel function to perform nonlinear mapping on the features, that is:

为非线性映射函数，则高炉历史数据与待测数据在核空间内在联合分布上的距为：in

is a nonlinear mapping function, then the distance between the blast furnace historical data and the data to be tested in the joint distribution in the kernel space is:

Introduce the kernel matrix K:

in:

Assuming that the required transformation matrix is W, the joint distribution distance between the blast furnace historical data and the data to be measured is:

Combined with the kernel matrix, the minimization problem of the above formula can be transformed into:

st W ^T KHK ^T W=I

Where: H=I _M+N -1/(M+N)11 ^T

The solution of this equation, the transformation matrix W, can be obtained by eigenvalue decomposition.

(2) The transformed feature variable is used as the input of the classifier. The connection weight between the feature variable and the classifier needs to be trained with the classification accuracy rate as the objective function. After convergence, the classification result of the classifier to be tested is the blast furnace fault. The categories are assigned to the data under test as new label values.

Fourth, iterative solution

Loop iterations are performed on steps 2 and 3, that is, the process of generating labels in weighted joint distribution adaptation, calculating weights and updating parameters is iteratively obtained to obtain fault diagnosis results.

5. Substitute actual industrial data for verification

We take the production data of the No. ² blast furnace with a volume of 2650m3 in an ironmaking plant in October 2020 as the historical data of the blast furnace, that is, the source domain data, and the blast furnace production data in December as the data to be measured, that is, the target domain data, which includes ' Oxygen enrichment rate', 'air permeability index', 'CO', 'H2', 'CO2', 'standard wind speed', 'oxygen enrichment flow', 'cold air flow', 'blast kinetic energy', 'both gas volume'','both gas index', 'theoretical combustion temperature', 'top pressure', 'oxygen-rich pressure', 'cold air pressure', 'total differential pressure', 'hot air pressure', 'actual wind speed', 'cold air temperature'','hot air temperature', 'top temperature northeast', 'top temperature southwest', 'top temperature northwest', 'top temperature southeast', 'top temperature downpipe', 'resistance coefficient', 'blast humidity', ' There are a total of 29 parameters in this hour's actual coal injection volume' and 'last hour's actual coal injection volume', with the same sampling rate. The validity of the model obtained by training is verified on the data to be tested, that is, the blast furnace production data in December.

Figures 2 and 3 respectively show the original distribution of the blast furnace data to be measured in December and the visualization results displayed by t-sne after the blast furnace to be measured data is classified into blast furnace faults by the method of the present invention. From the fault diagnosis results, it can be seen that the model works well. The classification effect is obvious, and it can accurately classify blast furnace failure samples, so it can be used in actual industrial production.

Claims (5)

1. a blast furnace fault diagnosis method based on weighted joint distribution adaptive neural network, is characterized in that, step is as follows: Step 1: Use the blast furnace historical data and blast furnace test data to train the weights of the deep neural network. The data value obtained by the last fully connected layer in the deep neural network is the extracted feature, and the fault diagnosis error of the blast furnace historical data is calculated. The sum of the distances from the extracted features is used as the loss function, and the weights are fixed after the training reaches a preset number of iterations or the loss function is lower than the preset value; Step 2: Use the fixed deep neural network in step 1 to perform the first feature extraction on the historical data of the blast furnace and the data to be measured, and generate a label value for the data to be measured in the blast furnace, that is, the historical data and the data to be measured are transferred. On the basis of using the blast furnace fault diagnosis knowledge learned from the blast furnace historical data and the data to be measured, a nonlinear mapping from blast furnace process variables to blast furnace fault categories is initially formed; Step 3: Based on the label value of the data to be tested, calculate the proportion of each fault category in the data to be tested and the historical data of the blast furnace respectively, and compare the ratio of the data to be tested of the blast furnace with the proportion of the corresponding category in the historical data to obtain the ratio. The prior distribution weight of the category of the blast furnace data to be tested is multiplied by the corresponding blast furnace historical data features extracted in step 2, and together with the characteristics of the blast furnace to be tested data extracted in step 2 to form a feature variable matrix, after weighting, The fault category distribution of the blast furnace historical data tends to be consistent with the fault category distribution of the blast furnace data to be tested, and the priori category distribution adaptation of the two is realized; Step 4: Introduce the kernel method, map the feature variables, obtain new feature variables, and transform the feature variables in the kernel space, so that the feature vectors extracted from the historical data of the blast furnace and the data to be tested are distributed in the marginal distribution and conditions. The sum of the distances on the distribution is the smallest. Therefore, the joint distribution adaptation of the blast furnace historical data and the data to be measured is realized by the method of the kernel method and the transformation matrix; Step 5: The transformed feature variable is used as the input of the classifier, and the connection weight between the feature variable and the classifier is trained with the classification accuracy rate as the objective function. After convergence, the classification result of the classifier to be tested is the blast furnace fault category. As a new label value assigned to the data to be tested; Step 6: Repeat steps 3 to 5 until the distance between the eigenvectors of the blast furnace historical data and the data to be tested on the joint distribution and the classification accuracy rate are both stable, then fix the model parameters, and perform the analysis on the data to be tested. Discriminate processing and generate fault diagnosis results. 2. method according to claim 1 is characterized in that, the structure of the described deep neural network of step 1 is as follows: deep neural network comprises three parts of input layer, hidden layer and output layer, and input layer is blast furnace process variable parameter The input layer includes air permeability index, cold air flow rate, hot air flow rate, top pressure, cold air pressure, and hot air pressure. The industrial process parameters that characterize the blast furnace production status, and the output layer is the blast furnace fault category layer, including difficult-to-run, suspended material, pipeline, and collapsed material. , furnace heat, furnace cooling and blast furnace faults related to blast furnace production process, the hidden layer is used to establish a non-linear mapping from blast furnace process variables to blast furnace fault categories, learn blast furnace fault diagnosis knowledge from blast furnace historical fault data, and establish blast furnace faults Diagnosis model; neurons in the same layer are not connected, neurons between layers are fully connected, and each connection has a weight to represent the strength of the connection between neurons; define neurons with a hidden layer greater than or equal to 2 The network is a deep neural network, and the mathematical model of a deep neural network is:

in,

is the output of the jth hidden layer unit of the i-th layer of the neural network, and denote h _i as the i-th layer of the neural network, then h ₀ is the input layer of the neural network, and h _k+1 is the output layer of the neural network; the value of j is based on the network The number of neurons in the i-th layer is determined, and the number of neurons in the i-th layer is denoted as _zi , then the value of each layer j is 1 to _zi ; W(i,j) is the j-th layer of the i-th layer The weight matrix corresponding to the neuron;

is the bias term corresponding to the jth neuron in the i-th layer, b _k+1 is the bias term corresponding to the output layer unit; y represents the output of the neural network, M is the total number of samples of the blast furnace historical data, and N is the blast furnace waiting time The total number of samples of the test data, f( ) and g( ) are the activation functions of the hidden layer unit and the output unit, respectively,

Represents the maximum value of the i-th sample in the neurons of the output layer, s _j represents the value of the j-th neuron in the output layer, extracts the data of the fully connected layer as a feature vector, and extracts the historical data of the blast furnace and the data to be tested. The feature vectors are recorded as x ^s and x ^t respectively, and the sum of the fault diagnosis error of the blast furnace historical data and the distance between the features extracted from the two sets of data is used as the loss function, which is the following formula (3):

3. method according to claim 1, is characterized in that, the step of weighting described in step 3 is as follows: record blast furnace fault altogether C type, c represents blast furnace fault type, when c takes the real number of 1 to C, represents corresponding The specific fault types, including pipeline, downward, difficult to travel, suspended material, M is the total number of samples of blast furnace historical data, and the number of samples belonging to the c-type fault type is M _C , correspondingly, N is the sample of blast furnace data to be measured. The total number, in which the number of samples belonging to the c-type fault type is N _C , the label value of the historical data is recorded as y ^s , the label value of the data to be measured is recorded as y ^t , and the distribution of blast furnace historical data and data to be measured is recorded as p _s respectively (·) and p _t (·), the proportions of various fault samples in the blast furnace historical data and the data to be measured are:

Then the corresponding weights of various fault data in the blast furnace historical data are:

After multiplying the weight by the historical data of the blast furnace, the prior distribution of the historical data of the blast furnace is:

Therefore, the fault category distribution of the blast furnace historical data and the category prior distribution of the blast furnace data to be tested tend to be consistent, and the prior category distribution of the two can be adapted. After weighting, the blast furnace historical data feature matrix X ^s and the test data The characteristic matrix X ^t constitutes the characteristic matrix X. 4. method according to claim 1, is characterized in that, the step of kernel method described in step 4 and joint distribution adaptation is as follows: choose kernel function such as Gaussian kernel function, carry out nonlinear mapping to feature, namely:

in

is a nonlinear mapping function, then in the kernel space, the distance from the joint distribution of the blast furnace historical data and the data to be measured is:

Introduce the kernel matrix K:

in:

Assuming that the required transformation matrix is W, the joint distribution distance between the blast furnace historical data and the data to be measured is:

Combined with the kernel matrix, the minimization problem of equation (14) can be transformed into:

st W ^T KHK ^T W=I (15) Where: H=I _M+N -1/(M+N)11 ^T (16)

The transformation matrix in the eigenspace can be solved by eigenvalue decomposition. 5. method according to claim 1, is characterized in that, the step of loop iteration described in step 6 is as follows: carry out iterative solution to step 3 to step 5, namely generate label in weighted joint distribution adaptation, calculate weight and The process of updating parameters is iterative to obtain fault diagnosis results.

Images