A semi-supervised soft-sensing method for coke oven flue temperature based on manifold regularization

Abstract

The present invention discloses a semi-supervised coke oven fire channel temperature soft measurement method based on manifold regularization, the method comprising: forming a labeled sample set with the top temperature and fire channel temperature of the collected time series matching, and using the top temperature collected at other times as an unlabeled sample set; determining the topological structure of the network; then performing multi-layer ELM-AE unsupervised feature extraction on the labeled sample set and the unlabeled sample set; using the extracted multi-layer ELM-AE unsupervised features as the input of the semi-supervised ELM regression layer, obtaining the output weight, and establishing a complete semi-supervised soft measurement model; collecting test input sample data and performing normalization processing, and then inputting it into the semi-supervised soft measurement model to obtain the output of the test sample. The present invention can realize the soft measurement of the fire channel temperature of the coke oven at low cost, which plays a key role in stabilizing coking production, improving product quality and reducing energy consumption.

Description

Technical Field

[0001] The invention relates to the field of soft measurement of coke oven flue temperature, and in particular to a semi-supervised soft measurement method of coke oven flue temperature based on manifold regularization. Background Art

[0002] Coke oven is one of the most important industrial furnaces in coal chemical industry and the key equipment for producing coke. Coke is widely used as raw material or fuel in blast furnace ironmaking, casting, calcium carbide, gasification and non-ferrous metal smelting. In the steel industry, the quality and output of coke directly affect the production quality and production cost of steel production. The temperature of coke oven is the most important process parameter affecting the quality of coke and the energy consumption of coking. It can ensure the quality and output of coke, so the detection and control of the temperature of coke oven is very important.

[0003] During the coke oven heating and combustion process, the most direct way to measure the temperature of the coke oven fire channel is to install a temperature measuring thermocouple in each combustion chamber, and use the thermocouple to directly measure the top temperature of a pair of combustion holes in each combustion chamber. This temperature measurement method can reflect the fire channel temperature more comprehensively and timely. However, the coke oven fire channel temperature is high and the environment is harsh. If the online detection instrument is directly installed, the equipment loss will be large. At the same time, due to the particularity of the coke oven structure, the installation and maintenance costs are too high.

[0004] The main detection method in the actual industrial process is that workers use infrared thermometers to measure manually through the temperature measuring holes on the furnace top at regular intervals. This is labor-intensive and has a long measurement cycle, far from meeting the requirements of real-time detection. Considering the difficulty of detection and cost control, the soft measurement method mines the data information of the coke oven heating and combustion process, establishes a mathematical model between the easily measured process variables (the top temperature of the regenerator) and the coke oven fire channel temperature, and uses auxiliary variables as model input to estimate the coke oven fire channel temperature, becoming an efficient and feasible method for coke oven fire channel temperature detection.

[0005] There are two problems in the soft measurement of the coke oven fire channel temperature. First, it is difficult to collect the coke oven fire channel temperature. According to the coke oven production process, the fire channel temperature is collected every four hours, and the coke oven temperature is obtained by calculating the average value; the detection frequency of the regenerator temperature is higher. Considering the switching operation, a regenerator temperature can be obtained every hour. It can be seen that the number of labeled training samples is very limited. Traditional soft measurement models often only use labeled samples and ignore unlabeled samples, resulting in a large amount of information contained in unlabeled samples being ignored. Second, the complex working conditions of the coke oven lead to a complex relationship between the input and output data. How to better mine and express the information between the data is very important for improving the accuracy of the model. Summary of the invention

[0006] The present invention provides a semi-supervised coke oven flue temperature soft measurement method based on manifold regularization. The method of realizing coke oven flue temperature soft measurement modeling at low cost will play a key role in stabilizing coking production, improving product quality and reducing energy consumption. In order to solve the technical problems mentioned in the above background technology, the present invention provides the following technical solutions:

[0007] The present invention provides a semi-supervised coke oven flue temperature soft measurement method based on manifold regularization, comprising:

[0008] Step S1, under normal operating conditions, collecting the top temperature of the regenerator during the heating and combustion process of the coke oven as a process variable, and the fire channel temperature as a target variable, and normalizing these data;

[0009] Step S2, the collected time-series matched storage roof temperature and fire channel temperature are used as a labeled sample set, and the storage roof temperature collected at other times is used as an unlabeled sample set;

[0010] Step S3, determining the topological structure of the network, including the number of hidden layers, the number of nodes and the activation function of the network;

[0011] Step S4, performing unsupervised feature extraction of the labeled sample set and the unlabeled sample set using a multi-layer extreme learning machine-based autoencoder (ELM-AE);

[0012] Step S5, using the multi-layer ELM-AE unsupervised features extracted in step S3 as the input of the semi-supervised ELM regression layer, obtaining the output weights, and establishing a complete semi-supervised soft measurement model;

[0013] Step S6: collect test input sample data and perform normalization processing, and then input it into the semi-supervised soft sensor model to obtain the output of the test sample.

[0014] Preferably, step S1 specifically comprises:

[0015] The top temperature of the regenerator during the heating and combustion process of the coke oven is collected as the process variable, and the fire channel temperature is used as the target variable. Then, the collected data of the top temperature and fire channel temperature of the coke oven are normalized. The formula is:

[0016]

[0017] Where x′ is the data after normalization, x is the original data collected, and x min is the minimum value in the original process variable data, x max It is the maximum value in the original process variable data.

[0018] Preferably, step S2 is specifically:

[0019] The labeled sample set is represented as Where l represents the number of labeled samples, x i =[x i,1 , xi,2 ,…,x i,6 ] T ∈R 6 Indicates the six selected storage top temperatures as the process variables of the model; y i ∈R indicates that the target variable of the soft sensor model is the furnace temperature; the unlabeled sample set is represented by Where u represents the number of unlabeled samples.

[0020] Preferably, the step S4 is specifically as follows:

[0021] Step 41: Randomly generate the input weights and bias parameters of each ELM-AE hidden layer and orthogonalize them to map the input data to different spatial dimensions. The expression is:

[0022]

[0023] Step 42: Take the coke oven top temperature as the input sample X as the target output X1=X of the first ELM-AE, calculate the output weight β1, and let The output matrix H1 of the first hidden layer is used as the input and target output of the next ELM-AE, and the training is carried out layer by layer to obtain the hidden layer weights of each layer; among which, the output H1 of the i-th hidden layer i Calculated as:

[0024] H i =g(H i-1 ·w i )

[0025] where w i (i=1,2,…,h) represents the weight between the i-1th hidden layer and the i-th hidden layer.

[0026] Preferably, step S5 is specifically as follows:

[0027] Step 51: Establish a similarity matrix S = [s ij ]∈R (l +u)*(l+u) ,s ij Represents sample x i and sample x j The similarity between them is calculated by Gaussian kernel function:

[0028]

[0029] Where N krepresents the set of k neighbors of sample x, σ represents the kernel parameter;

[0030] Step 52: Calculate the Laplacian matrix L=DS of the graph containing labeled samples and unlabeled samples, where is a similar diagonal matrix of samples, and the value of the diagonal element is the sum of all the values ​​in each column;

[0031] Step 53: Add the manifold regularization term about the similarity between the labeled samples and the unlabeled samples to the objective function of ELM as a constraint, combine the labeled samples and the unlabeled samples to find the distribution characteristics, and use the local structure information between the coke oven top temperature data to make up for the missing coke oven fire channel temperature information, and finally obtain the objective function of the semi-supervised ELM:

[0032]

[0033] where ‖β‖ 2 is the regularization term, ‖Y-Gα‖ 2 is the empirical risk error, β T H T LHβ is the manifold regularization term, the balance parameter λ is set to control the geometry of the data distribution and prevent overfitting, Y represents the actual output of the labeled sample, G∈R l×n and H∈R (l×n)×n Hidden layer neurons corresponding to labeled samples and all samples respectively;

[0034] Step 54: Depending on the number of samples, there are two different solutions for the output weight. When the number of labeled samples is greater than the number of hidden layer nodes, the solution for the output weight is:

[0035] β * =(I k +CG T G+λH T LH) -1 CG T Y

[0036] When the number of labeled samples is less than the number of hidden layer nodes, the solution of the output weight is:

[0037]

[0038] Among them I k and I l+u are k-dimensional and l+u-dimensional identity matrices respectively, and Represents an augmented matrix, with the first l rows containing G and Y and the last u rows containing 0.

[0039] The beneficial effects of the present invention are as follows: taking into account the complex relationship between input and output data due to the complex working conditions of the coke oven, through the unsupervised feature extraction of multi-layer ELM-AE, deeper features of the data are mined; in addition, in order to utilize a large amount of information contained in the unlabeled samples, a manifold regularization term about the similarity between the labeled samples and the unlabeled samples is added to the loss function of the ELM regression layer, the labeled samples and the unlabeled samples are combined to find the distribution characteristics, and the local structural information between the coke oven top temperature data is used to make up for the missing fire channel temperature information; the method of realizing the soft measurement modeling of the coke oven fire channel temperature at a low cost will play a key role in stabilizing the coking production, improving product quality and reducing energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is the temperature curve of the heat storage chamber;

[0041] Figure 2 This is the flow chart of the soft measurement method for the temperature of the coke oven flue;

[0042] Figure 3 This is a schematic diagram of the timing matching of the coke oven fire channel temperature input and output. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0044] Example 1

[0045] The coke oven is composed of many combustion chambers and carbonization chambers arranged alternately, and each carbonization chamber corresponds to a heat storage chamber. Figure 1 As shown. After being preheated in the regenerator, the heating gas enters the combustion chamber for combustion, and the heat generated is transferred to the adjacent carbonization chamber by radiation. The coal in the carbonization chamber is dry distilled under air-tight conditions to form the final product, coke. Therefore, the coke oven is composed of many combustion subsystems connected in series, and each subsystem affects each other. The coke oven flue temperature reflects the average flue temperature of each combustion subsystem, and the flue temperature of each combustion subsystem will affect the coke oven flue temperature.

[0046] Refer to the attached Figure 2 , the following are the detailed steps of the present invention for realizing a semi-supervised coke oven flue temperature soft measurement method based on manifold regularization.

[0047] Step S1, under normal operating conditions, collecting the top temperature of the regenerator during the heating and combustion process of the coke oven as a process variable, and the fire channel temperature as a target variable, and normalizing these data;

[0048] Specifically, during the coke oven heating and combustion process, thermocouples are installed on the top of some regenerators to collect the top temperature of the regenerator once a minute. In the actual operation process, the top temperature changes as follows: Figure 1 As shown, the reversing operation is performed every half an hour. From 0 to 30 minutes, the high-temperature exhaust gas generated in the combustion chamber enters the heat storage chamber, causing the temperature of the heat storage chamber to rise; after reversing, 30 to 60 minutes, the gas and air involved in the combustion enter the heat storage chamber for preheating, and the temperature of the heat storage chamber decreases. In this process, the temperature of the exhaust gas generated by the combustion in the combustion chamber has the greatest correlation with the fire channel temperature. After the exhaust gas temperature enters the heat storage chamber, the temperature of the heat storage chamber continues to rise. At about 20 minutes, the temperature gradually stabilizes, as shown in Figure 1 Therefore, the temperature of point A 20 minutes after the exhaust gas enters the regenerator is selected as the top temperature of the regenerator. A top temperature can be extracted every hour. Considering the characteristics of temperature change, the value of 10-minute smoothing filter is taken as the top temperature of the regenerator. For example: Figure 1 At point A, the temperature collected every 15-25 minutes is smoothed and filtered as the top storage temperature. The cycle of the fire channel temperature detection is 4 hours, and the fire channel temperature is collected every 4 hours.

[0049] Then the collected data of coke oven top temperature and fire channel temperature are normalized, and the normalization formula is:

[0050]

[0051] Where x′ is the data after normalization, x is the original data collected, and x min is the minimum value in the original process variable data, x max It is the maximum value in the original process variable data.

[0052] Step S2, the collected time-series matched storage roof temperature and fire channel temperature are used as a labeled sample set, and the storage roof temperature collected at other times is used as an unlabeled sample set;

[0053] Specifically, the set of labeled samples is represented as Where l represents the number of labeled samples, x i =[x i,1 ,x i,2 ,…,x i,6 ] T ∈R 6 Indicates the six selected storage top temperatures as the process variables of the model; y i∈R indicates that the target variable of the soft sensor model is the furnace temperature; the unlabeled sample set is represented by Where u represents the number of unlabeled samples.

[0054] Step S3, determining the topological structure of the network;

[0055] Specifically, the number of hidden layers of the network is 2, the number of nodes is set to [30, 20], and the activation function is the Sigmoid function.

[0056] Step S4, performing unsupervised feature extraction of the labeled sample set and the unlabeled sample set using a multi-layer extreme learning machine-based autoencoder (ELM-AE);

[0057] First, the input weights and bias parameters of each ELM-AE hidden layer are randomly generated and orthogonalized to map the input data X to different spatial dimensions. The expression is:

[0058]

[0059] Take the input sample X as the target output X1=X of the first ELM-AE and find the output weight β1. The output matrix H1 of the first hidden layer is used as the input and target output of the next ELM-AE, and the training is carried out layer by layer to obtain the hidden layer weights of each layer; among which, the output H1 of the i-th hidden layer i Calculated as:

[0060] H i =g(H i-1 ·w i )

[0061] Among them, g(·) represents the activation function, w i (i=1,2,…,h) represents the weight between the i-1th hidden layer and the i-th hidden layer.

[0062] Step S5, using the multi-layer ELM-AE unsupervised features extracted in step S4 as the input of the semi-supervised ELM regression layer, obtaining the output weights, and establishing a complete semi-supervised soft measurement model;

[0063] First, determine the topology of the extreme learning machine. ELM is a single hidden layer feedforward network with the characteristics of fast training speed and good generalization performance. Assume that the network contains T hidden layer nodes, expressed as:

[0064]

[0065] where a i is the weight vector between the i-th hidden layer node and the input node, b i is the bias of the i-th node, β iis the weight vector connecting the i-th hidden layer node and the output node.

[0066] The calculation expression of the hidden layer output matrix H is:

[0067]

[0068] The target variable matrix Y is:

[0069]

[0070] The goal of ELM is to find the minimum training error and the minimum output weight norm, that is,

[0071]

[0072] Where C is the penalty coefficient.

[0073] Solving the output weights using the least squares regression algorithm yields:

[0074]

[0075] in is the Moore–Penrose generalized inverse of the hidden layer output matrix H.

[0076] In order to utilize the large amount of information contained in unlabeled samples, a manifold regularization term about the similarity between labeled samples and unlabeled samples is added to the loss function of ELM, and the labeled samples and unlabeled samples are combined to find the distribution characteristics.

[0077] Specifically, firstly, a similarity matrix S = [s ij ]∈R (l+u)*(l+u) ,s ij Represents sample x i and sample x j The similarity between them can be calculated by Gaussian kernel function:

[0078]

[0079] Where N k represents the set of k neighbors of sample x, and σ represents the kernel parameter.

[0080] The expression of the manifold regularization function is defined as follows:

[0081]

[0082] Where L=DS is the Laplacian matrix of the graph containing labeled and unlabeled samples. is a similar diagonal matrix of samples, where the diagonal elements are the sum of all values ​​in each column.

[0083] Finally, the manifold regularization function is added to the objective function of ELM as a constraint, and the local structure information between the coke oven top temperature data is used to make up for the missing fire channel temperature information. The objective function of the semi-supervised ELM regression layer is as follows:

[0084]

[0085]

[0086]

[0087] Substituting the constraints into the objective function, we can obtain:

[0088]

[0089] where ‖β‖ 2 is the regularization term, ‖Y-Gα‖ 2 is the empirical risk error, β T H T LHβ is the manifold regularization term, and the balance parameter λ is set to control the geometry of the data distribution and prevent overfitting. Y represents the actual output of the labeled sample, G∈R l×n and H∈R (l×n)×n The hidden layer neurons corresponding to the labeled samples and all samples respectively.

[0090] Depending on the number of samples, there are two different solutions for the output weight. If the number of labeled samples is greater than the number of hidden layer nodes, the solution for the output weight is:

[0091] β * =(I k +CG T G+λH T LH) -1 CG T Y

[0092] When the number of labeled samples is less than the number of hidden layer nodes, the solution of the output weight is:

[0093]

[0094] Among them I k and I l+u are k-dimensional and l+u-dimensional identity matrices respectively, and Represents an augmented matrix, with the first l rows containing G and Y and the last u rows containing 0.

[0095] Step S6: collect test input sample data and perform normalization processing, and then input it into the semi-supervised soft sensor model to obtain the output of the test sample.

[0096] The present invention is not limited to the embodiments described above. The above description of the specific embodiments is intended to describe and illustrate the technical solution of the present invention. The above specific embodiments are merely illustrative and not restrictive. Without departing from the scope of the present invention and the scope of protection of the claims, a person of ordinary skill in the art can also make many forms of specific changes under the guidance of the present invention, which all fall within the scope of protection of the present invention.

Claims

1. A semi-supervised coke oven flue temperature soft-sensing method based on manifold regularization, characterized in that: include: Step S1, under normal operating conditions, collecting the top temperature of the regenerator during the heating and combustion process of the coke oven as a process variable, and the fire channel temperature as a target variable, and normalizing these data; Step S2, the collected time-series matched storage roof temperature and fire channel temperature are used as a labeled sample set, and the storage roof temperature collected at other times is used as an unlabeled sample set; Step S3, determining the topological structure of the network, including the number of hidden layers, the number of nodes and the activation function of the network; Step S4, performing multi-layer ELM-AE unsupervised feature extraction on the labeled sample set and the unlabeled sample set; Step S5, using the multi-layer ELM-AE unsupervised features extracted in step S4 as the input of the semi-supervised ELM regression layer, obtaining the output weights, and establishing a complete semi-supervised soft measurement model; Step S6, collecting test input sample data and normalizing it, and then inputting it into the semi-supervised soft sensor model to obtain the output of the test sample; Step S5 is specifically as follows: Step 51: Establish a similarity matrix S = [s ij ]∈R (l+u)*(l+u) ,s ij Represents sample x i and sample x j The similarity between them is calculated by Gaussian kernel function: Where N k represents the set of k neighbors of sample x, σ represents the kernel parameter; Step 52: Calculate the Laplacian matrix L=DS of the graph containing labeled samples and unlabeled samples, where is a similar diagonal matrix of samples, and the value of the diagonal element is the sum of all the values ​​in each column; Step 53: Add the manifold regularization term about the similarity between the labeled samples and the unlabeled samples to the objective function of ELM as a constraint, combine the labeled samples and the unlabeled samples to find the distribution characteristics, and use the local structure information between the coke oven top temperature data to make up for the missing fire channel temperature information, and finally obtain the objective function of the semi-supervised ELM: where ||β|| 2 is the regularization term, ||Y-Gβ|| 2 is the empirical risk error, β T H T LHβ is the manifold regularization term, and the balance parameter λ It is the geometric shape setting that controls the data distribution and can prevent overfitting. Y represents the actual output of the fire channel temperature, G∈R l×n and H∈R (l×n)×n Hidden layer neurons corresponding to labeled samples and all samples respectively; Step 54: Depending on the number of samples, there are two different solutions for the output weight. When the number of labeled samples is greater than the number of hidden layer nodes, the solution for the output weight is: β * =(I k +CG T G+λH T LH) -1 CG T Y When the number of labeled samples is less than the number of hidden layer nodes, the solution of the output weight is: Among them I k and I l+u are k-dimensional and l+u-dimensional identity matrices respectively, and Represents an augmented matrix, with the first l rows containing G and Y and the last u rows containing 0.

2. The semi-supervised coke oven flue temperature soft sensing method based on manifold regularization as claimed in claim 1, characterized in that: Step S2 is specifically as follows: The labeled sample set is represented as Where l represents the number of labeled samples, x i =[x i,1 ,x i,2 ,…,x i,6 ] T ∈R 6 Indicates the six selected storage top temperatures as the process variables of the model; y i ∈R indicates that the target variable of the soft sensor model is the furnace temperature; the unlabeled sample set is represented by Where u represents the number of unlabeled samples.

3. The semi-supervised coke oven flue temperature soft sensing method based on manifold regularization as claimed in claim 1, characterized in that: The step S4 is specifically as follows: Step 41: Randomly generate the input weights and bias parameters of each ELM-AE hidden layer and orthogonalize them. The expression is: Step 42: Take the coke oven top temperature as input data X as the target output X1=X of the first ELM-AE, calculate the output weight β1, and let The output matrix H1 of the first hidden layer is used as the input and target output of the next ELM-AE, and the training is carried out layer by layer to obtain the hidden layer weights of each layer; among which, the output H1 of the i-th hidden layer i Calculated as: H i =g(H i-1 ·w i ) where w i (i=1,2,…,h) represents the weight between the i-1th hidden layer and the i-th hidden layer.

Images