System and method for on-line detection, regulation and control of indexes of salt-making mother liquor

Abstract

The method comprises the following steps: establishing a near infrared spectrum database of key indexes of the salt-making mother liquor, and carrying out chemical component calibration by combining a laboratory standard method; introducing the mother liquor into an optical detection pool through a bypass circulation system, and collecting near infrared spectrum data; constructing a semi-supervised regression model between the spectral data and the target parameters, and calculating the sulfate radical content, the chloride ion content and the solid-to-liquid ratio in the mother liquor by utilizing the established semi-supervised regression model; a detection result is transmitted to a control system, and crystallization process parameters are adjusted. According to the scheme provided by the invention, the continuous and real-time monitoring of the salt-making mother liquor is realized through the combination of the bypass circulation system and the near infrared spectrum technology; the semi-supervised regression model is utilized, the flow model depth reversible network and the maximum mutual information coefficient are combined, overlapping signals and complex background interference in spectral data are effectively decoupled, and the accuracy and stability of multi-component quantitative analysis are improved.

Description

Technical Field

[0001] This application relates to the field of salt production mother liquor detection technology, and in particular to an online detection and control system and method for salt production mother liquor indicators. Background Technology

[0002] The composition and physicochemical properties of the mother liquor used in salt production, such as sulfate content, chloride ion content, conductivity, and temperature, directly impact the quality of the final product and production efficiency. Most manufacturers currently employ conventional chemical analysis methods for analysis and monitoring, involving manual sampling, complex pretreatment processes, and laboratory testing. This results in very long sample analysis and result feedback cycles, sometimes lasting several hours, which fails to meet the demands of modern production for efficiency, real-time monitoring, and automation. Furthermore, the lag in detection can hinder timely adjustments to production processes, leading to batch defects or food safety risks. With the development of intelligent manufacturing and digital technologies, online detection technologies, such as near-infrared spectroscopy, Raman spectroscopy, machine vision, and electrochemical sensors, are widely used in the food industry. These technologies enable real-time, non-destructive, and continuous monitoring of key indicators, significantly reducing detection time and improving data reliability, thereby enhancing the control precision and response speed of the production process.

[0003] Vacuum salt production equipment and thermal nitrate extraction-salt-nitrate co-production process utilize the principle that salt and mirabilite have different solubilities at different temperatures. Through continuous multi-effect evaporation and crystallization, both salt and mirabilite products are produced simultaneously. The content of various chemical components in the mother liquor is one of the important indicators of the operational reliability of the vacuum salt production equipment. It not only affects the degree of separation between salt and mirabilite products but also plays a decisive role in the quality of the final product and production efficiency. In particular, the sulfate content in the mother liquor is crucial for ensuring the quality of the salt-nitrate products. However, the spectral signals detected by line detection technology overlap significantly, information is intertwined and coupled, and a large amount of redundant noise affects the accuracy of multi-component quantitative analysis of the spectrum.

[0004] Therefore, it is necessary to design an online detection and control system and method that combines an online monitoring system with a vacuum salt production device and a thermal nitrate extraction-salt-nitrate co-production process, so as to achieve rapid, efficient and accurate detection of the content of various chemical components in the mother liquor, and improve the quality and stability of salt-nitrate products. Summary of the Invention

[0005] To address or partially address the problems existing in related technologies, this application provides an online detection and control system and method for salt production mother liquor indicators, aiming to solve the technical problems of severe spectral signal overlap, information intertwining and coupling, and a large amount of redundant noise in the online detection of salt production mother liquor.

[0006] This application provides a method for online detection and control of indicators in salt production mother liquor, including: Establish a near-infrared spectral database of key indicators of salt production mother liquor, and determine the chemical composition in combination with laboratory standard methods; The mother liquor is introduced into the optical detection cell through a bypass circulation system to collect near-infrared spectral data. A semi-supervised regression model was constructed between spectral data and target parameters. This model was used to calculate the sulfate content, chloride ion content, and solid-liquid ratio in the mother liquor. Specifically, the high-dimensional original spectral data was mapped to the latent space of latent variables using a deep reversible network of the flow model, and then decoupled into independent latent variables that satisfy a Gaussian distribution. Based on the maximum mutual information coefficient, the latent variable with the highest information entropy of the component data was selected to effectively mine key component information features and eliminate complex background interference in the spectral data. A partial least squares regression model was constructed using the latent variable representation after deep spectral decoupling to quantitatively analyze the sulfate content, chloride ion content, and solid-liquid ratio. The test results are transmitted to the control system to adjust the crystallization process parameters.

[0007] Optionally, in some implementations, a semi-supervised regression model is constructed between the spectral data and the target parameters, including: Acquire spectral data of salt production mother liquor from the near-infrared spectral database, normalize the spectral data, and divide the dataset. A deep invertible network based on a flow model is constructed, containing four additively coupled layers. The maximum number of iterations of the adaptive moment estimation optimizer is set, the activation function is a linear rectified function, and the loss function is: (14) By adding the maximum information coefficient to the deep invertible network of the flow model, a semi-supervised model is formed. The semi-supervised model network was pre-trained using the partitioned dataset to learn the distribution of the original spectral features; The learned original spectral data distribution is transformed into latent variables that follow a Gaussian normal distribution of the same dimension, and the maximum information coefficient value of each latent variable and the predicted value is calculated. By using the greedy forward selection method, the set of variables with the largest mutual information values ​​is selected, and a partial least squares regression model is established.

[0008] Optionally, in some implementations, the deep reversible network of the flow model includes: Due to severe overlap of spectral features and significant mutual interference, multi-component spectra are often not simple linear superpositions of single-component spectral signals. To effectively extract the spectral features of each component and achieve feature decoupling, a transformation function is required. The original spectral signal X is transformed into a latent variable space Z that satisfies a simpler distribution than the original spectral signal. The transformation function is... A mapping from x to z can be established for the distribution. Each point in the distribution can be found in the distribution. Find the corresponding point; To achieve high-quality spectral data representation and deeply mine key information in the latent variable space, a reversible transformable flow model deep network is used to fit the transformation function. The flow model reversible transformable deep network structure contains four additive coupling layers. Since the transformed latent features are independent in each dimension and each feature has its own independent intrinsic meaning, the transformed latent features are easier to learn. Because the flow model deep reversible network structure is reversible, it ensures no information loss during variable transformation, retains the effective information in the original spectral data, and better enables feature decoupling representation learning. When the variable transformation is smooth and invertible, mutual information can remain unchanged because of the transformation function. If it is reversible, then: (1) (2) In the formula, g is the inverse function of the transformation function f, f is the transformation function that maps the original spectrum to the latent variable space, and z is the latent variable constructed by the mapping function; According to the variable transformation theorem, we have (3) To theoretically achieve lossless decoupling and reversibility of information, the dimensions of x and z should be consistent, and the distributions of the independent latent variables should satisfy a standard Gaussian distribution. For the original spectral signal distribution, we have... (4) In the formula, It can be any conditional Gaussian distribution or Dirac distribution; To solve for the parameters The maximum likelihood method and the objective function are employed. Determine the nonlinear transformation ; Since the transformation function f is invertible, the data dimensions of x and z are both D-dimensional, and It follows a standard Gaussian distribution, therefore we have (5) choose Dirac distribution Since g is a bijective function, we can obtain the following from the above equation: (6) In the formula, yes The Jacobian matrix of x; The objective function to be optimized is: (7) To solve the optimization objective, a network model is trained to approximate the fitted transformation function. Due to the complexity of Jacobi determinant calculation, selecting a suitable network model to simplify the calculation process is crucial. The nonlinear independent component estimation model is a deep learning framework based on the normalized flow model, which uses affine coupling transformation to achieve decoupling and representation of the data; Nonlinear independent component estimation models have three important model structures: an additive coupling layer, a shuffling layer, and a scaling layer. The additive coupling layer divides the original data into two equal parts. and And perform different operations respectively, as shown below: , (8) , (9) In the formula, m represents any complex function, and a multilayer perceptron (MLP) is used to fit the function; and These are newly generated potential variables; All Jacobian matrices based on additive coupling layer transformations have a triangular determinant with all diagonals being 1, and can be represented by block matrices: (10) In the formula, These are the first d elements of sample X. These are elements between d+1 and D; It is an identity matrix of size 1; Therefore, it can be concluded that the determinant det can be guaranteed in each calculation. The value is 1, which simplifies the calculation of the objective function and makes the Jacobian determinant of the transformation function easier to calculate; Considering that the transformation capability of a single additive coupling layer is weak and it is difficult to map complex nonlinear relationships, the nonlinear independent component estimation model consists of four additive coupling layers. Before each additive coupling, the order of dimensions is shuffled or reversed to fully mix the information and achieve stronger nonlinearity. (11) Furthermore, the nonlinear independent component estimation model incorporates a scaling layer at the end to compress the manifold, resulting in: (12) In the formula, It is a tensor product, and S is a diagonal scaling matrix that can identify the importance of features; Therefore, the Jacobian matrix of the scaling transformation layer is: (13) The determinant value of the scaling layer is Combining the optimization objective function, the final optimization objective function of the nonlinear independent component estimation model is: (14) In summary, the nonlinear independent component estimation model can solve the problem of the difficulty in calculating the Jacobian determinant during lossless feature extraction.

[0009] Optionally, in some implementations, the maximum information coefficient value for each latent variable and predicted value is calculated, including: Since the latent space dimension of the flow model is large, the maximum information coefficient is used to eliminate feature redundancy in order to improve the interpretability of independent latent variables. Assumption It is a variable. Here, n is the number of samples, and the mutual information (MI) between the variable and the target attribute is: (15) In the formula, and These are the marginal probability densities of Z and Y. It is the joint probability density of Z and Y; Assumption It is a finite binary dataset. The variable Z is divided into z intervals, forming a z×y grid G. Mutual information is calculated in each grid, and the maximum mutual information of the set D is... (16) In the formula, The mutual information (MI) value under the probability distribution is divided by the grid of the dataset; The maximum MI value across all partitions is regularized to form a feature matrix. Its definition is: (17) The maximum information coefficient (MIC) is defined as follows: (18) In the formula, This represents the upper bound of the grid, z×y; Through standardization The larger the MIC value, the more relevant the features are; by performing discrete optimization on continuous variables with unequal intervals, the correlation of variables can be mined using the MIC value.

[0010] Optionally, in some implementations, a partial least squares regression model is established, including: Let the largest latent variable representation matrix be The response variable matrix is ; Data standardization, in order to eliminate the influence of units, firstly involves... and The sample was standardized, with a mean of 0 and a variance of 1. Extracting the first latent variable, since we need to consider both... and ,let Each variable and To perform regression, use the regression coefficients as weights to calculate... A linear combination of these, the weight vector is denoted as... ; Using weight vector calculate The first latent variable : (19) use To each and Perform regression to obtain the load vector and scalar : (20) (twenty one) In the formula, and It is the residual matrix; Remove the interpreted information and use the residual matrix. and Replace the original and , , ; Repeated iterations, for new and Repeat the steps of extracting the first latent variable to extract the second latent variable. , ..., each iteration can extract information related to the response variable matrix from the residual; The iteration stops when the number of latent variables extracted reaches a preset value, which is determined by cross-validation, or when the residual is sufficiently small. Ultimately, a regression model is obtained for the original variables, where all latent variables t are original variables. A linear combination of these ultimately writes the model back to its original form. In the form of , B is the regression coefficient matrix and F is the residual matrix.

[0011] Optionally, in some embodiments, a near-infrared spectral database of key indicators of the brine mother liquor is established, including: Collect laboratory standard spectral data of historical salt production mother liquor, and determine the corresponding chemical composition values ​​of the salt production mother liquor using current standard methods, as the calibration basis data for online analysis; Collect spectra of materials with the same chemical composition as the main component of the salt production mother liquor; A database was established using laboratory standard spectral data, calibration baseline data values, and material spectra as data sources.

[0012] Optionally, in some implementations, near-infrared spectral data is acquired, including: The mother liquor is introduced into the optical detection cell through a bypass circulation system, and the characteristic absorption spectra of the components in the salt production mother liquor to near-infrared light are detected by a Fourier transform near-infrared spectrometer, with the near-infrared light wavelength range of 780-2526nm.

[0013] The second aspect of this application provides an online detection and control system for indicators of salt production mother liquor, including: The near-infrared spectroscopy detection module includes an optical detection cell, a spectrometer, an optical fiber, and a lens; The optical detection cell is made of stainless steel with a Teflon coating on the inner wall for corrosion and adhesion prevention. Its spectral range covers 780–2526 nm, making it suitable for high-salt, high-temperature mother liquor environments; The data processing module is used to carry semi-supervised regression models; it integrates Sunny Lib software for real-time data processing and model calibration; and it has database storage, self-learning, and adaptive calibration functions. Process control module: Transmits detection data to the DCS / PLC system in real time via Modbus-TCP protocol, and dynamically adjusts evaporation and crystallization process parameters based on the detection results; The auxiliary system, including filters, valves, and bypass circulation lines, ensures continuous and stable sample flow; its explosion-proof design makes it suitable for industrial environments.

[0014] The technical solution provided in this application may include the following beneficial effects: By combining a bypass circulation system with near-infrared spectroscopy, manual sampling and laboratory analysis are eliminated, shortening the detection cycle and enabling continuous, real-time monitoring of key indicators in salt production mother liquor. A semi-supervised regression model, combined with a deep reversible network of the flow model and the maximum mutual information coefficient, effectively decouples overlapping signals and complex background interference in spectral data, uncovers key component information features, and improves the accuracy and stability of multi-component quantitative analysis. The detection results are transmitted to the control system in real time, dynamically adjusting evaporation and crystallization process parameters to achieve closed-loop control of the production process, improving the quality consistency and production efficiency of salt and nitrate products, and reducing the risk of batch non-compliance due to detection delays.

[0015] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0016] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0017] Figure 1 This is a schematic flowchart illustrating the online detection and control method for salt production mother liquor indicators in the embodiments of this application; Figure 2 This is a schematic diagram of the reversible transformation flow model deep network structure of the online detection and control method for salt production mother liquor indicators shown in the embodiments of this application; Figure 3 This is a flowchart illustrating the semi-supervised regression model of the online detection and control method for salt production mother liquor indicators shown in the embodiments of this application. Detailed Implementation

[0018] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0019] Vacuum salt production equipment and thermal nitrate extraction-salt-nitrate co-production process utilize the principle that salt and mirabilite have different solubilities at different temperatures. Through continuous multi-effect evaporation and crystallization, both salt and mirabilite products are produced simultaneously. The content of various chemical components in the mother liquor is one of the important indicators of the operational reliability of the vacuum salt production equipment. It not only affects the degree of separation between salt and mirabilite products but also plays a decisive role in the quality of the final product and production efficiency. In particular, the sulfate content in the mother liquor is crucial for ensuring the quality of the salt-nitrate products. However, the spectral signals detected by line detection technology overlap significantly, information is intertwined and coupled, and a large amount of redundant noise affects the accuracy of multi-component quantitative analysis of the spectrum.

[0020] To address the aforementioned issues, this application provides an online detection and control system and method for salt production mother liquor indicators. This system combines a bypass circulation system with near-infrared spectroscopy, eliminating the need for manual sampling and laboratory analysis, shortening the detection cycle, and enabling continuous, real-time monitoring of key indicators in the salt production mother liquor. Utilizing a semi-supervised regression model, combined with a deep reversible network of the flow model and the maximum mutual information coefficient, it effectively decouples overlapping signals and complex background interference in the spectral data, uncovers key component information features, and improves the accuracy and stability of multi-component quantitative analysis. The detection results are transmitted to the control system in real time, dynamically adjusting the evaporation and crystallization process parameters to achieve closed-loop control of the production process, improving the quality consistency and production efficiency of salt and nitrate products, and reducing the risk of batch non-compliance due to detection delays.

[0021] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0022] Figure 1 This is a schematic flowchart illustrating the online detection and control method for salt production mother liquor indicators as shown in the embodiments of this application.

[0023] See Figure 1 A method for online detection and control of indicators in salt production mother liquor, comprising: S101. Establish a near-infrared spectral database of key indicators of salt production mother liquor, and determine the chemical composition in combination with laboratory standard methods; Specifically, a near-infrared spectral database of key indicators for salt production mother liquor will be established, including: Collect laboratory standard spectral data of historical salt production mother liquor, and determine the corresponding chemical composition values ​​of the salt production mother liquor using current standard methods, as the calibration basis data for online analysis; Collect spectra of materials with the same chemical composition as the main component of the salt production mother liquor; A database was established using laboratory standard spectral data, calibration baseline data values, and material spectra as data sources.

[0024] S102. The mother liquor is introduced into the optical detection cell through the bypass circulation system to collect near-infrared spectral data. Specifically, near-infrared spectral data is collected, including: The mother liquor is introduced into the optical detection cell through a bypass circulation system, and the characteristic absorption spectra of the components in the salt production mother liquor to near-infrared light are detected by a Fourier transform near-infrared spectrometer, with the near-infrared light wavelength range of 780-2526nm.

[0025] S103. Construct a semi-supervised regression model between spectral data and target parameters, and use the established semi-supervised regression model to calculate the sulfate content, chloride content, and solid-liquid ratio in the mother liquor. Specifically, a semi-supervised regression model is constructed between spectral data and target parameters, including: Acquire spectral data of salt production mother liquor from the near-infrared spectral database, normalize the spectral data, and divide the dataset. A deep invertible network based on a flow model is constructed, consisting of four additively coupled layers. The maximum number of iterations of the adaptive moment estimation optimizer is set, and the activation function is a linear rectified function. By adding the maximum information coefficient to the deep invertible network of the flow model, a semi-supervised model is formed. The semi-supervised model network was pre-trained using the partitioned dataset to learn the distribution of the original spectral features; The learned original spectral data distribution is transformed into latent variables that follow a Gaussian normal distribution of the same dimension, and the maximum information coefficient value of each latent variable and the predicted value is calculated. By using the greedy forward selection method, the set of variables with the largest mutual information values ​​is selected, and a partial least squares regression model is established.

[0026] Specifically, deep reversible networks for flow models include: Due to severe overlap of spectral features and significant mutual interference, multi-component spectra are often not simple linear superpositions of single-component spectral signals. To effectively extract the spectral features of each component and achieve feature decoupling, a transformation function is required. The original spectral signal X is transformed into a latent variable space Z that satisfies a simpler distribution than the original spectral signal. The transformation function is... A mapping from x to z can be established for the distribution. Each point in the distribution can be found in the distribution. Find the corresponding point; To achieve high-quality spectral data representation and deeply mine key information in the latent variable space, a reversible transformable flow model deep network is used to fit the transformation function. The flow model reversible transformable deep network structure contains four additive coupling layers. Since the transformed latent features are independent in each dimension and each feature has its own independent intrinsic meaning, the transformed latent features are easier to learn. Because the flow model deep reversible network structure is reversible, it ensures no information loss during variable transformation, retains the effective information in the original spectral data, and better enables feature decoupling representation learning. When the variable transformation is smooth and invertible, mutual information can remain unchanged because of the transformation function. If it is reversible, then: (1) (2) In the formula, g is the inverse function of the transformation function f, f is the transformation function that maps the original spectrum to the latent variable space, and z is the latent variable constructed by the mapping function; According to the variable transformation theorem, we have (3) To theoretically achieve lossless decoupling and reversibility of information, the dimensions of x and z should be consistent, and the distributions of the independent latent variables should satisfy a standard Gaussian distribution. For the original spectral signal distribution, we have... (4) In the formula, It can be any conditional Gaussian distribution or Dirac distribution; To solve for the parameters The maximum likelihood method and the objective function are employed. Determine the nonlinear transformation ; Since the transformation function f is invertible, the data dimensions of x and z are both D-dimensional, and It follows a standard Gaussian distribution, therefore we have (5) choose Dirac distribution Since g is a bijective function, we can obtain the following from the above equation: (6) In the formula, yes The Jacobian matrix of x; The objective function to be optimized is: (7) To solve the optimization objective, a network model is trained to approximate the fitted transformation function. Due to the complexity of Jacobi determinant calculation, selecting a suitable network model to simplify the calculation process is crucial. The nonlinear independent component estimation model is a deep learning framework based on the normalized flow model, which uses affine coupling transformation to achieve decoupling and representation of the data; Nonlinear independent component estimation models have three important model structures: an additive coupling layer, a shuffling layer, and a scaling layer. The additive coupling layer divides the original data into two equal parts. and And perform different operations respectively, as shown below: , (8) , (9) In the formula, m represents any complex function, and a multilayer perceptron (MLP) is used to fit the function; and These are newly generated potential variables; All Jacobian matrices based on additive coupling layer transformations have a triangular determinant with all diagonals being 1, and can be represented by block matrices: (10) In the formula, These are the first d elements of sample X. These are elements between d+1 and D; It is an identity matrix of size 1; Therefore, it can be concluded that the determinant det can be guaranteed in each calculation. The value is 1, which simplifies the calculation of the objective function and makes the Jacobian determinant of the transformation function easier to calculate; Considering that the transformation capability of a single additive coupling layer is weak and it is difficult to map complex nonlinear relationships, the nonlinear independent component estimation model consists of four additive coupling layers. Before each additive coupling, the order of dimensions is shuffled or reversed to fully mix the information and achieve stronger nonlinearity. (11) Furthermore, the nonlinear independent component estimation model incorporates a scaling layer at the end to compress the manifold, resulting in: (12) In the formula, It is a tensor product, and S is a diagonal scaling matrix that can identify the importance of features; Therefore, the Jacobian matrix of the scaling transformation layer is: (13) The determinant value of the scaling layer is Combining the optimization objective function, the final optimization objective function of the nonlinear independent component estimation model is: (14) In summary, the nonlinear independent component estimation model can solve the problem of the difficulty in calculating the Jacobian determinant during lossless feature extraction.

[0027] Specifically, calculate the maximum information coefficient value for each latent variable and predicted value, including: Since the latent space dimension of the flow model is large, the maximum information coefficient is used to eliminate feature redundancy in order to improve the interpretability of independent latent variables. Assumption It is a variable. Here, n is the number of samples, and the mutual information (MI) between the variable and the target attribute is: (15) In the formula, and These are the marginal probability densities of Z and Y. It is the joint probability density of Z and Y; Assumption It is a finite binary dataset. The variable Z is divided into z intervals, forming a z×y grid G. Mutual information is calculated in each grid, and the maximum mutual information of the set D is... (16) In the formula, The mutual information (MI) value under the probability distribution is divided by the grid of the dataset; The maximum MI value across all partitions is regularized to form a feature matrix. Its definition is: (17) The maximum information coefficient (MIC) is defined as follows: (18) In the formula, This represents the upper bound of the grid, z×y; Through standardization The larger the MIC value, the more relevant the features are; by performing discrete optimization on continuous variables with unequal intervals, the correlation of variables can be mined using the MIC value.

[0028] Specifically, establishing a partial least squares regression model includes: Let the largest latent variable representation matrix be The response variable matrix is ; Data standardization, in order to eliminate the influence of units, firstly involves... and The sample was standardized, with a mean of 0 and a variance of 1. Extracting the first latent variable, since we need to consider both... and ,let Each variable and To perform regression, use the regression coefficients as weights to calculate... A linear combination of these, the weight vector is denoted as... ; Using weight vector calculate The first latent variable : (19) use To each and Perform regression to obtain the load vector and scalar : (20) (twenty one) In the formula, and It is the residual matrix; Remove the interpreted information and use the residual matrix. and Replace the original and , , ; Repeated iterations, for new and Repeat the steps of extracting the first latent variable to extract the second latent variable. , ..., each iteration can extract information related to the response variable matrix from the residual; The iteration stops when the number of latent variables extracted reaches a preset value, which is determined by cross-validation, or when the residual is sufficiently small. Ultimately, a regression model is obtained for the original variables, where all latent variables t are original variables. A linear combination of these ultimately writes the model back to its original form. In the form of , B is the regression coefficient matrix and F is the residual matrix.

[0029] S104. Transmit the detection results to the control system and adjust the crystallization process parameters.

[0030] Corresponding to the aforementioned application function implementation device embodiments, this application also provides an online detection and control system for salt production mother liquor indicators and corresponding embodiments.

[0031] An online detection and control system for indicators of salt production mother liquor, comprising: The near-infrared spectroscopy detection module includes an optical detection cell, a spectrometer, an optical fiber, and a lens; The optical detection cell is made of stainless steel with a Teflon coating on the inner wall for corrosion and adhesion prevention. Its spectral range covers 780–2526 nm, making it suitable for high-salt, high-temperature mother liquor environments; The data processing module is used to carry semi-supervised regression models; it integrates Sunny Lib software for real-time data processing and model calibration; and it has database storage, self-learning, and adaptive calibration functions. The process control module transmits the detection data to the DCS / PLC system in real time via the Modbus-TCP protocol and dynamically adjusts the evaporation and crystallization process parameters based on the detection results. The auxiliary system, including filters, valves, and bypass circulation lines, ensures continuous and stable sample flow; its explosion-proof design makes it suitable for industrial environments.

[0032] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for online detection and control of indicators in salt production mother liquor, characterized in that, include: Establish a near-infrared spectral database of key indicators of salt production mother liquor, and determine the chemical composition in combination with laboratory standard methods; The mother liquor is introduced into the optical detection cell through a bypass circulation system to collect near-infrared spectral data. A semi-supervised regression model was constructed between spectral data and target parameters. This model was used to calculate the sulfate content, chloride ion content, and solid-liquid ratio in the mother liquor. Specifically, the high-dimensional original spectral data was mapped to the latent space of latent variables using a deep reversible network of the flow model, and then decoupled into independent latent variables that satisfy a Gaussian distribution. Based on the maximum mutual information coefficient, the latent variable with the highest information entropy of the component data was selected to effectively mine key component information features and eliminate complex background interference in the spectral data. A partial least squares regression model was constructed using the latent variable representation after deep spectral decoupling to quantitatively analyze the sulfate content, chloride ion content, and solid-liquid ratio. The test results are transmitted to the control system to adjust the crystallization process parameters.

2. The online detection and control method for salt production mother liquor indicators according to claim 1, characterized in that, The construction of the semi-supervised regression model between spectral data and target parameters includes: Obtain the spectral data of the salt production mother liquor from the near-infrared spectral database, normalize the spectral data, and divide the dataset. A deep invertible network based on a flow model is constructed, containing four additively coupled layers. The maximum number of iterations of the adaptive moment estimation optimizer is set, the activation function is a linear rectified function, and the loss function is: （14） By adding the maximum information coefficient to the deep invertible network of the flow model, a semi-supervised model is formed. The semi-supervised model network was pre-trained using the partitioned dataset to learn the distribution of the original spectral features; The learned original spectral data distribution is transformed into latent variables that follow a Gaussian normal distribution of the same dimension, and the maximum information coefficient value of each latent variable and the predicted value is calculated. By using the greedy forward selection method, the set of variables with the largest mutual information values ​​is selected, and a partial least squares regression model is established.

3. The online detection and control method for salt production mother liquor indicators according to claim 2, characterized in that, The deep reversible network of the flow model includes: By transforming the function The original spectral signal X is transformed into a latent variable space Z that satisfies a simpler distribution than the original spectral signal. The transformation function is... A mapping from x to z can be established for the distribution. Each point in the distribution can be found in the distribution. Find the corresponding point; A deep network with reversible transformation of the flow model is used to fit the transformation function. The deep network structure with reversible transformation of the flow model contains four additive coupling layers. The reversible flow model deep network adopts a nonlinear independent component estimation model. The nonlinear independent component estimation model is a deep learning framework based on the normalized flow model. It uses affine coupling transformation to achieve decoupling and representation of data. The nonlinear independent component estimation model has three important model structures: additive coupling layer, order shuffling layer, and scaling transformation layer. The final objective function of the nonlinear independent component estimation model is: （14） Nonlinear independent component estimation models can solve the problem of the difficulty in calculating the Jacobian determinant during lossless feature extraction.

4. The online detection and control method for salt production mother liquor indicators according to claim 2, characterized in that, The calculation of the maximum information coefficient value for each latent variable and predicted value includes: Since the latent space dimension of the flow model is large, the maximum information coefficient is used to eliminate feature redundancy in order to improve the interpretability of independent latent variables. Assumption It is a variable. Here, n is the number of samples, and the mutual information (MI) between the variable and the target attribute is: （15） In the formula, and These are the marginal probability densities of Z and Y. It is the joint probability density of Z and Y; Assumption It is a finite binary dataset; the variable Z is divided into z intervals, forming a z×y grid G. Mutual information is calculated in each grid, and the maximum mutual information of the set D is: （16） In the formula, The mutual information (MI) value under the probability distribution is divided by the grid of the dataset; The maximum MI value across all partitions is regularized to form a feature matrix. Its definition is: （17） The definition of the maximum information coefficient (MIC) is: （18） In the formula, This represents the upper bound of the grid, z×y; Through standardization The larger the MIC value, the more relevant the features are; by performing discrete optimization on continuous variables with unequal intervals, the correlation of variables can be mined using the MIC value.

5. The online detection and control method for salt production mother liquor indicators according to claim 2, characterized in that, The establishment of the partial least squares regression model includes: Let the largest latent variable representation matrix be The response variable matrix is ; Data standardization, in order to eliminate the influence of units, firstly involves... and The sample is standardized, with a mean of 0 and a variance of 1. Extracting the first latent variable, since we need to consider both... and ,let Each variable and To perform regression, use the regression coefficients as weights to calculate... A linear combination of these, the weight vector is denoted as... ; Using weight vector calculate The first latent variable : （19） use To each and Perform regression to obtain the load vector and scalar : （20） （21） In the formula, and It is the residual matrix; Remove the interpreted information and use the residual matrix. and Replace the original and , , ; Repeated iterations, for new and Repeat the steps of extracting the first latent variable to extract the second latent variable. , ..., each iteration can extract information related to the response variable matrix from the residuals; The iteration stops when the number of latent variables extracted reaches a preset value, which is determined by cross-validation, or when the residual is sufficiently small. Ultimately, a regression model is obtained for the original variables, where all latent variables t are original variables. A linear combination of these ultimately writes the model back to its original form. In the form of , B is the regression coefficient matrix and F is the residual matrix.

6. The online detection and control method for salt production mother liquor indicators according to claim 1, characterized in that, The establishment of a near-infrared spectral database of key indicators for salt production mother liquor includes: Collect laboratory standard spectral data of historical salt production mother liquor, and determine the corresponding chemical composition values ​​of the salt production mother liquor using current standard methods, as the calibration basis data for online analysis; Collect spectra of materials with the same chemical composition as the main component of the salt production mother liquor; A database was established using laboratory standard spectral data, calibration baseline data values, and material spectra as data sources.

7. The online detection and control method for salt production mother liquor indicators according to claim 1, characterized in that, The acquired near-infrared spectral data includes: The mother liquor is introduced into the optical detection cell through a bypass circulation system, and the characteristic absorption spectra of the components in the salt production mother liquor to near-infrared light are detected by a Fourier transform near-infrared spectrometer, with the near-infrared light wavelength range of 780-2526nm.

8. An online detection and control system for indicators of brine mother liquor, used to execute the online detection and control method for indicators of brine mother liquor according to any one of claims 1-7, characterized in that, include: The near-infrared spectroscopy detection module includes an optical detection cell, a spectrometer, an optical fiber, and a lens; The optical detection cell is made of stainless steel with a Teflon coating on the inner wall for corrosion and adhesion prevention. Its spectral range covers 780–2526 nm, making it suitable for high-salt, high-temperature mother liquor environments; The data processing module is used to carry semi-supervised regression models; it integrates Sunny Lib software for real-time data processing and model calibration; and it has database storage, self-learning, and adaptive calibration functions. The process control module transmits the detection data to the DCS / PLC system in real time via the Modbus-TCP protocol and dynamically adjusts the evaporation and crystallization process parameters based on the detection results. The auxiliary system, including filters, valves, and bypass circulation lines, ensures continuous and stable sample flow; its explosion-proof design makes it suitable for industrial environments.

Images