Sulfur recovery device soft measurement method based on non-linear semi-supervised regression model

Abstract

The invention discloses a sulfur recovery device soft measurement method based on a nonlinear semi-supervised regression model. According to the method, a radial basis function and kernel transformation are introduced into a linear model structure, and then a data-driven soft measurement model in a nonlinear semi-supervised form is provided by fusing a large number of label-free samples through a Laplace regularization technology and is used for online detection of the content of hydrogen sulfide and sulfur dioxide in a sulfur recovery device. Compared with a traditional linear regression model, the non-linear semi-supervised regression model can more effectively utilize a large amount of label-free data, meanwhile, the non-linear relation between the input data and the output data is established, and therefore the application effect of the soft measurement model is improved.

Description

Technical Field

[0001] The invention belongs to the field of soft measurement modeling and application of chemical processes, and in particular relates to a soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model. Background Art

[0002] Sulfur recovery device is an important chemical industry production process. In this device, how to obtain the content values ​​of hydrogen sulfide and sulfur dioxide in the tail gas of the device in real time is a very important indicator for the quality monitoring and optimization control of the sulfur recovery device, which directly affects the economic benefits and indicators of the entire device. However, for the key quality variables of the device: the measurement of the content values ​​of hydrogen sulfide and sulfur dioxide, there is currently a lack of direct and effective online measurement means. It is usually necessary to use expensive analytical instruments or manual testing methods to indirectly obtain their measured values. In recent years, data-driven soft measurement technology has provided a new means for estimating variables that are difficult to measure in chemical processes, and the content values ​​of hydrogen sulfide and sulfur dioxide are estimated in real time through other easily measured variables in the device. Linear regression is a very commonly used soft measurement model and has been widely used. However, due to the linear constraints of the model and the number of modeling samples, its application effect is greatly affected, and it is difficult to meet the needs of practical applications in many occasions. Based on the architecture of the basic linear regression model, the present invention introduces a nonlinear radial basis function, and generalizes the linear regression model to a nonlinear form through the kernel function technique. At the same time, Laplace regularization is introduced into the optimization function of the model to integrate a large amount of unlabeled sample data, which effectively alleviates the overfitting risk of the nonlinear regression model and improves its soft measurement estimation effect in the sulfur recovery unit. Summary of the invention

[0003] The purpose of the present invention is to provide a soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model in view of the shortcomings of the soft measurement performance of key variables in the sulfur recovery device.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] A soft sensing method for a sulfur recovery unit based on a nonlinear semi-supervised regression model comprises the following steps:

[0006] (1) Use conventional measuring instruments to collect variable data in the sulfur recovery unit, mainly including easily measurable variables such as temperature, pressure, and flow rate, analyze the correlation between these measured variables and the target variables (i.e., hydrogen peroxide and sulfur dioxide content), select appropriate variables and data samples, and form a training data sample set for modeling: X∈R N×m , where N is the number of sample data sets and m is the number of process variables.

[0007] (2) In order to obtain the content data of two target variables: hydrogen oxide and sulfur dioxide, we measure them through analytical instruments, and calibrate them through on-site manual sampling and offline laboratory analysis to ensure the accuracy of target variable measurement. Compared with other easy-to-measure process variables, the two target variables are relatively difficult to obtain, so the number of samples collected will be much smaller, denoted as Y∈R n , as the soft sensor model outputs the training sample set, where n is the number of samples in the data set.

[0008] (3) Each data variable collected in the sulfur recovery unit is preprocessed to remove obvious outliers and erroneous data, and the mean of each process variable and target variable is made zero and the variance is 1 through normalization method, so as to obtain a standardized training data set under a unified dimensional scale.

[0009] (4) After dimension scale standardization, the input and output data are paired according to the time axis and divided into a labeled data set {X∈Rn×m,Y∈Rn×r} and an unlabeled data set X∈R (N-n)×m ,Due to the different difficulty levels of variable data collection methods, ,in general, the capacity of unlabeled datasets is much larger than that of labeled datasets.

[0010] (5) Based on a small-scale labeled dataset {X∈R n×m ,Y∈R n×r}, based on the linear regression model, a nonlinear radial basis function is introduced, and the regression coefficient matrix of the model is obtained using the least squares optimization method.

[0011] (6) Based on the basic nonlinear regression model, the model parameters are regularized to alleviate the risk of model overfitting caused by the small number of training data samples.

[0012] (7) In the nonlinear regularized regression model, the Laplace regularization term is introduced to integrate a large number of unlabeled data samples X∈R (N-n)×m , a better description of the input data distribution of the soft measurement model is given, and a semi-supervised nonlinear regression model is established to improve the performance of online soft measurement.

[0013] (8) New data samples are collected from the sulfur recovery unit, normalized, and the dimensions of each variable are unified. The new data are then input into a nonlinear semi-supervised regression model to calculate the online estimated values ​​of hydrogen sulfide and sulfur dioxide content corresponding to the new data.

[0014] An electronic device comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the program, the soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model is implemented.

[0015] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model.

[0016] Beneficial effects of the present invention:

[0017] The present invention establishes a regression relationship between conventional process variables that are easy to measure in a sulfur recovery device and two key variables (hydrogen sulfide and sulfur dioxide content) on the device, constructs a nonlinear semi-supervised soft measurement modeling, and performs online estimation of two gas content values ​​that are difficult to measure through the variables that are easy to measure in the device, thereby providing important support for the quality control of the sulfur recovery device. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 :Flow chart of sulfur recovery unit. DETAILED DESCRIPTION

[0019] The present invention aims at the problem of online real-time detection of hydrogen sulfide and sulfur dioxide gas contents in a sulfur recovery device, and realizes online soft measurement of two important gas contents by using a nonlinear semi-supervised regression model through conventional process variables that are easy to measure in the device.

[0020] Embodiment: A soft sensing method for a sulfur recovery device based on a nonlinear semi-supervised regression model, the method is as follows:

[0021] The first step is to use conventional measuring instruments to collect variable data in the sulfur recovery unit, mainly including easily measurable variables such as temperature, pressure, and flow, analyze the correlation between these measured variables and the target variables (i.e., hydrogen peroxide and sulfur dioxide content), select appropriate variables and data samples, and form a training data sample set for modeling: X∈R N×m , where N is the number of sample data sets and m is the number of process variables.

[0022] In the second step, in order to obtain the content data of two target variables: hydrogen oxide and sulfur dioxide, we measured them through analytical instruments, and calibrated them through on-site manual sampling and offline laboratory analysis to ensure the accuracy of target variable measurement. Compared with other easy-to-measure process variables, the two target variables are relatively difficult to obtain, so the number of samples collected will be much smaller, denoted as Y∈R n, as the soft sensor model outputs the training sample set, where n is the number of samples in the data set.

[0023] The third step is to preprocess each data variable collected in the sulfur recovery unit to remove obvious outliers and erroneous data, and use the normalization method to make the mean of each process variable and target variable zero and the variance 1, so as to obtain a standardized training data set under a unified dimensional scale.

[0024] The fourth step is to pair the input and output data after dimensional scale standardization according to the time axis and divide them into labeled data sets {X∈Rn×m,Y∈Rn×r} and unlabeled data sets X∈R (N-n)×m ,Due to the different difficulty levels of variable data collection methods, ,in general, the capacity of unlabeled datasets is much larger than that of labeled datasets.

[0025] The fifth step is based on a small-scale labeled dataset {X∈R n×m ,Y∈R n×r}, based on the linear regression model, a nonlinear radial basis function is introduced, and the regression coefficient matrix of the model is obtained using the least squares optimization method:

[0026]

[0027] in, is the nonlinear radial basis function, α is the Lagrange multiplier of the optimization model, W ls is the least squares optimization coefficient matrix of the nonlinear regression model. From the above formula, we can see that from the linear regression model to the nonlinear regression model, the optimization goal has actually changed from the coefficient matrix W ls Transformed to α.

[0028] The sixth step is to regularize the model parameters based on the basic nonlinear regression model to alleviate the risk of model overfitting caused by small training data samples. The input data and output data are re-expressed as X = [x1, x2, ..., x N ],Y=[y1,y2,...,y n ,0,...,0], and define J=diag{1,...,1,0,...,0} as a diagonal matrix. The optimization objectives and parameters of the model are as follows:

[0029]

[0030] Among them, λ is the regularization parameter, is a kernel function, which can usually be selected as a Gaussian kernel. Compared with the basic nonlinear regression model, the regularized nonlinear regression model can alleviate the risk of model overfitting caused by small number of modeling samples to a certain extent.

[0031] The seventh step is to introduce the Laplace regularization term into the nonlinear regularized regression model and integrate a large number of unlabeled data samples X∈R (N-n)×m , better describe the input data distribution of the soft sensor model, establish a semi-supervised nonlinear regression model, and improve the performance of online soft sensors. The final optimization objective function and parameters of the model are as follows:

[0032]

[0033] Where β is the Laplace regularization parameter, L = DM, M is the weight matrix in the adjacency graph of the input variable data, defined as follows:.

[0034]

[0035] Where E is the edge set containing adjacent vertex pairs (x i ,x j ).

[0036] The eighth step is to collect new data samples from the sulfur recovery unit, normalize them, unify the dimensions of each variable, and then input the new data into the nonlinear semi-supervised regression model to calculate the online estimated values ​​of hydrogen sulfide and sulfur dioxide content corresponding to the new data:

[0037]

[0038] If we use sampling and actual calibration to determine the quality index value y new , we can get the estimation error of the nonlinear regression soft sensor model as:

[0039]

[0040] In addition, the root mean square error indicator can usually be used to evaluate the performance of the soft sensor model, which is defined as follows:

[0041]

[0042] Where T is the number of test samples of the soft sensor model.

[0043] The effectiveness of the present invention is illustrated below by combining a specific practical application case of a sulfur recovery device. Figure 1As shown, for the production process of the device, we collected a total of 1,000 input variable data with good quality. However, due to the difficulty in measuring the target variable, we can only collect part of the output variable data. Therefore, except for a part of the input variable data that is marked, most of the input variable data is unmarked. In order to establish a nonlinear semi-supervised regression soft measurement model, we selected a total of 5 process variables that are highly correlated with the target variable as the input variables of the soft measurement model through mechanism analysis and consulting the experience of engineers. These variables are MEA gas flow, primary airflow, secondary airflow, SWS area gas flow and airflow in the SWS area. Next, the implementation steps of the present invention are described in detail in conjunction with the device:

[0044] 1. In the first step, the 1000 collected modeling samples are divided into labeled data sets and unlabeled data sets, and each variable is normalized so that the mean of each input variable and output variable of the soft sensor model is zero and the variance is 1, and a new standardized modeling data sample set is obtained.

[0045] 2. Nonlinear semi-supervised regression soft sensor modeling

[0046] Five process variables selected from the sulfur recovery unit are used as the input of the soft sensor model, and the contents of hydrogen sulfide and sulfur dioxide are used as the output of the soft sensor model. According to the detailed implementation steps given in the method of the present invention, an online soft sensor model based on nonlinear semi-supervised regression is established.

[0047] 3. Acquire new data information in the production process of the sulfur recovery unit in real time, first normalize it using model parameters, and then input it into the nonlinear semi-supervised soft sensor regression model to calculate the hydrogen sulfide and sulfur dioxide content values ​​of the new data in real time.

[0048] In order to test the effectiveness of the new invention method, we collected 1000 new data samples in real time during the operation of the sulfur recovery unit. First, we used the normalization parameters of the soft measurement model to process them so that the dimension of the new data was consistent with the modeling data sample. Then, we used the established nonlinear semi-supervised soft measurement regression model to estimate each new data online, and compared the performance with several basic data-driven soft measurement models. The results are shown in Table 1. The table lists the performance of each soft measurement model under different proportions of unlabeled data samples. It can be seen that under different conditions, the method of the present invention has achieved the best estimation effect, that is, the root mean square error RMSE of the target variable estimation is the smallest.

[0049] Table 2: Online estimation performance of various soft sensor models

[0050]

[0051] The above embodiments are used to illustrate the present invention rather than to limit the present invention. Any modification and change made to the present invention within the spirit of the present invention and the protection scope of the claims shall fall within the protection scope of the present invention.

Claims

1. A soft sensing method for a sulfur recovery unit based on a nonlinear semi-supervised regression model, characterized in that: The method comprises the following steps: (1) Use conventional measuring instruments to collect variable data in the sulfur recovery unit, mainly including temperature, pressure, and flow, which are easy to measure. Analyze the correlation between these measured variables and the target variables (i.e., hydrogen peroxide and sulfur dioxide content), select appropriate variables and data samples, and form a training data sample set for modeling: X∈R N×m , where N is the number of sample data sets, m is the number of process variables, (2) In order to obtain the content data of two target variables: hydrogen oxide and sulfur dioxide, on the one hand, they are measured by analytical instruments, and on the other hand, they are corrected by on-site manual sampling and offline laboratory analysis to ensure the accuracy of the target variable measurement. The number of samples collected for the two target variables will be much smaller, denoted as Y∈R n , as the soft-sensing model output training sample set, where n is the number of samples in the data set, (3) Preprocess each data variable collected in the sulfur recovery unit to remove obvious outliers and erroneous data, and use the normalization method to make the mean of each process variable and target variable zero and the variance 1, so as to obtain a standardized training data set under a unified dimensional scale. (4) After dimension scale standardization, the input and output data are paired according to the time axis and divided into labeled data sets {X∈R n×m ,Y∈R n×r }, and unlabeled dataset X∈R (N-n)×m , (5) Based on a small-scale labeled dataset {X∈R n×m ,Y∈R n×r }, based on the linear regression model, the nonlinear radial basis function is introduced, and the regression coefficient matrix of the model is obtained by using the least squares optimization method. (6) Based on the basic nonlinear regression model, the model parameters are regularized to alleviate the risk of model overfitting caused by the small number of training data samples. (7) In the nonlinear regularized regression model, the Laplace regularization term is introduced to integrate a large number of unlabeled data samples X∈R (N-n)×m , better describe the input data distribution of the soft sensor model, establish a semi-supervised nonlinear regression model, and improve the performance of online soft sensors. (8) New data samples are collected from the sulfur recovery unit, normalized, and the dimensions of each variable are unified. The new data are then input into a nonlinear semi-supervised regression model to calculate the online estimated values ​​of hydrogen sulfide and sulfur dioxide content corresponding to the new data.

2. According to claim 1, a soft sensing method for sulfur recovery unit based on nonlinear semi-supervised regression model is characterized in that: The step (5) is specifically as follows: based on a small-scale labeled dataset {X∈R n×m ,Y∈R n×r }, based on the linear regression model, a nonlinear radial basis function is introduced, and the regression coefficient matrix of the model is obtained using the least squares optimization method: in, is the nonlinear radial basis function, α is the Lagrange multiplier of the optimization model, W ls is the least squares optimization coefficient matrix of the nonlinear regression model. From the above formula, we can see that from the linear regression model to the nonlinear regression model, the optimization goal has actually changed from the coefficient matrix W ls Transformed to α.

3. According to claim 1, a soft sensing method for sulfur recovery unit based on nonlinear semi-supervised regression model, characterized in that: The step (6) is specifically as follows: on the basis of the basic nonlinear regression model, the parameters of the model are regularized to alleviate the risk of model overfitting caused by the small number of training data samples, and the input data and output data are re-expressed as X = [x1, x2, ..., x N ],Y=[y1,y2,...,y n ,0,...,0], and define J=diag{1,...,1,0,...,0} as a diagonal matrix. The optimization objectives and parameters of the model are as follows: Among them, λ is the regularization parameter, is the kernel function, usually in the form of a Gaussian kernel.

4. According to claim 1, a soft sensing method for sulfur recovery device based on nonlinear semi-supervised regression model is characterized in that: The step (7) is specifically as follows: In the nonlinear regularized regression model, a Laplace regularization term is introduced to fuse a large number of unlabeled data samples X∈R (N-n)×m , to better describe the input data distribution of the soft sensor model, establish a semi-supervised nonlinear regression model, and improve the performance of online soft sensors. The final optimization objective function and parameters of the model are as follows: Where β is the Laplace regularization parameter, L = DM, M is the weight matrix in the adjacency graph of the input variable data, defined as follows:. Where E is the edge set containing adjacent vertex pairs (x i ,x j ), exp represents the operator of the exponential function.

5. The soft sensing method for sulfur recovery unit based on nonlinear semi-supervised regression model according to claim 1, characterized in that: The step (8) specifically comprises: collecting new data samples from the sulfur recovery device, normalizing the data samples, unifying the dimensions of various variables, and then inputting the new data into a nonlinear semi-supervised regression model to calculate the online estimated values ​​of hydrogen sulfide and sulfur dioxide content corresponding to the new data: If the quality index value obtained through sampling and actual calibration is y new , the estimation error of the nonlinear regression soft sensor model is obtained as: In addition, the root mean square error indicator can usually be used to evaluate the performance of the soft sensor model, which is defined as follows: Where T is the number of test samples of the soft sensor model.

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model as described in any one of claims 1 to 5 is implemented.

7. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by the processor, the soft measurement method for a sulfur recovery device based on a nonlinear semi-supervised regression model as described in any one of claims 1 to 5 is implemented.

Images