A soft-sensing method and device based on a supervised multi-latent variable structure

Abstract

The embodiment of the application discloses a kind of soft measurement method and device based on supervised multiple latent variable structure, which method comprises: the process data obtained in the process of thermal power generation production scheduling is input into the quality extraction model established in advance;Extract the quality related information of fuel by the quality extraction model;The quality related information includes process variable and quality variable;Wherein, the quality extraction model is a mathematical prediction model of supervised multiple latent variable structure, and the multiple latent variable includes: dynamic latent variable and static latent variable, the dynamic latent variable is used to extract the cross-correlation of the process variable and the quality variable, and the static latent variable is used to preserve the cross-correlation of the process variable and the quality variable.By the embodiment scheme, the real-time measurement of quality variable is realized.

Description

Technical Field

[0001] This application relates to fuel quality measurement technology, and more particularly to a soft measurement method and apparatus based on a supervised multi-potential variable structure. Background Technology

[0002] With the expansion of modern industrial processes such as power generation, the demands for measurement and control in large thermal power generating units are also increasing. In particular, the emphasis on energy conservation and efficiency has placed increasingly higher demands on the optimized and economical operation of these units. The simple detection and control of temperature and pressure in the past is far from sufficient to meet today's needs. Many parameters in the production scheduling process of thermal power plants, such as the oxygen content in flue gas, the carbon content in slag, and the excess oxygen content in the boiler, are of great significance to improving production efficiency.

[0003] However, due to technological and economic limitations, these crucial parameters are often difficult or impossible to detect using traditional hardware measurement sensors. Furthermore, many hardware measurement devices suffer from significant time lags, leading to delays in process control and impacting production. Some key process variables require offline chemical analysis, further exacerbating the time lag. For instance, measuring coal composition in thermal power generation requires complex chemical experiments to obtain results. To overcome these challenges, researchers have proposed soft measurement techniques that establish mathematical prediction models for key process variables based on process characteristics and data.

[0004] In soft sensor modeling, due to the complex experimental analysis and high labor costs required to obtain quality variables, the missing value rate often exceeds 80% during actual data collection. The missing quality variable values ​​actually exist, but current equipment or technology cannot obtain them in a timely manner. Therefore, the cross-correlation between variables and the autocorrelation between samples are still included in the dataset. Since quality variables are randomly missing, arbitrary data deletion is not allowed, as this would disrupt the original autocorrelation relationships between the data.

[0005] Cross-correlation, autocorrelation, and uncertainty (measurement noise, missing variables) are the main characteristics of thermal power generation scheduling data, and current technologies mostly address specific problems. However, in most cases, these data characteristics coexist in a single thermal power generation scheduling process, making it difficult for current soft measurement models to simultaneously account for all of these issues within a single thermal power generation scheduling process. Summary of the Invention

[0006] This application provides a soft measurement method and apparatus based on a supervised multi-latent variable structure, which can realize real-time measurement of quality variables.

[0007] This application provides a soft measurement method based on a supervised multi-latent variable structure, the method including:

[0008] The process data obtained during the thermal power generation scheduling process is input into a pre-established quality extraction model;

[0009] The quality-related information of fuel is extracted using the quality extraction model; the quality-related information includes process variables and quality variables.

[0010] The quality extraction model is a supervised mathematical prediction model with a multi-latent variable structure. The multi-latent variables include dynamic latent variables and static latent variables. The dynamic latent variables are used to extract the cross-correlation between the process variables and the quality variables, and the static latent variables are used to retain the cross-correlation between the process variables and the quality variables.

[0011] In an exemplary embodiment of this application, creating the quality extraction model may include:

[0012] Create a model structure with multiple latent variables;

[0013] The model structure is trained to obtain the quality extraction model.

[0014] In an exemplary embodiment of this application, creating a model structure with multiple latent variables may include:

[0015] Within the Gaussian framework, the dynamic latent variables are determined by a first-order Markov chain; the static latent variables are independent of each other.

[0016] The dynamic latent variables and the static latent variables are linearly combined to form the process variables and the quality variables.

[0017] In an exemplary embodiment of this application, training the model structure to obtain the quality extraction model may include:

[0018] The quality variables that require soft measurement during the thermal power generation scheduling process and the process variables related to the quality variables that require soft measurement are collected as the training dataset for the model structure.

[0019] Based on the training dataset, the log-likelihood function of the model structure is maximized using the Expectation Maximization (EM) algorithm to solve for the parameters of the model structure.

[0020] Substitute the solved parameters into the model structure to obtain the quality extraction model.

[0021] In an exemplary embodiment of this application, before solving for the parameters of the model structure by maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm based on the training dataset, the method may further include:

[0022] Determine whether there are missing quality variables in the training dataset;

[0023] If the training dataset has missing quality variables, then based on the semi-supervised algorithm, the Expectation Maximization (EM) algorithm is used to maximize the log-likelihood function of the model structure according to the training dataset, and the parameters of the model structure are solved.

[0024] If the training dataset does not have missing quality variables, then based on the fully supervised algorithm, the Expectation-Maximization (EM) algorithm is used to maximize the log-likelihood function of the model structure according to the training dataset, and the parameters of the model structure are solved.

[0025] In an exemplary embodiment of this application, the step of maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm based on the training dataset to solve for the parameters of the model structure may include: iteratively executing the following steps until a preset iteration stopping condition is met to obtain the updated parameters as the final parameters of the model structure:

[0026] The posterior probability distributions of the dynamic latent variables and the static latent variables are calculated using the E-step, and the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic latent variables and the static latent variables is derived.

[0027] Based on the posterior probability distribution and its form, the parameter update of the model structure is completed in M ​​steps.

[0028] In an exemplary embodiment of this application, the step of calculating the posterior probability distributions of the dynamic latent variables and the static latent variables using the E-step, and deriving the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic latent variables and the static latent variables, may include:

[0029] During the state prediction phase, based on the dynamic latent variable q from the previous time t-1... t-1 The posterior probability distribution is used to calculate the dynamic latent variable q at the current time t. t The posterior probability distribution;

[0030] Obtain the dynamic latent variable q at the current time t t and static latent variable h t The joint distribution representation of is used as the form of the posterior probability distribution;

[0031] During the observation correction phase, the dynamic latent variable q is adjusted based on the given observation process variables and observation quality variables. t and static latent variable h t The joint distribution representation is corrected;

[0032] According to the modified dynamic latent variable q t and static latent variable h t The joint distribution represents the acquisition of the dynamic latent variable q respectively. t and the static latent variable h t The marginal distribution of is used as the posterior probability distribution of the dynamic latent variables input in the next state prediction stage.

[0033] In an exemplary embodiment of this application, the step of updating the parameters of the model structure in M ​​steps based on the posterior probability distribution and the form of the posterior probability distribution may include:

[0034] In the M-step, the partial derivatives of the modified posterior probability distribution and its form are calculated, and the partial derivatives are set to zero to calculate the parameters of the model structure.

[0035] In an exemplary embodiment of this application, the extraction of fuel mass-related information through the mass extraction model may include:

[0036] Calculate the dynamic and static latent variables corresponding to each process data based on the parameters of the quality extraction model;

[0037] The predicted value of the quality variable of the process data is calculated based on the posterior probability distribution of the dynamic and static latent variables corresponding to each process data.

[0038] This application also provides a soft measurement device based on a supervised multi-potential variable structure, which may include a processor and a computer-readable storage medium. The computer-readable storage medium stores instructions that, when executed by the processor, implement the soft measurement method.

[0039] Compared with related technologies, the embodiments of this application may include: inputting process data obtained during the thermal power generation scheduling process into a pre-established quality extraction model; extracting fuel quality-related information through the quality extraction model; the quality-related information includes process variables and quality variables; wherein, the quality extraction model is a supervised mathematical prediction model with a multi-latent variable structure, the multi-latent variables include: dynamic latent variables and static latent variables, the dynamic latent variables are used to extract the cross-correlation relationship between the process variables and the quality variables, and the static latent variables are used to retain the cross-correlation relationship between the process variables and the quality variables. Through this embodiment, real-time measurement of quality variables is achieved.

[0040] Other features and advantages of this application will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the application. Other advantages of this application can be realized and obtained by means of the solutions described in the description and the accompanying drawings. Attached Figure Description

[0041] The accompanying drawings are used to provide an understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0042] Figure 1 This is a flowchart of a soft measurement method based on a supervised multi-latent variable structure, as described in an embodiment of this application.

[0043] Figure 2 This is a schematic diagram of the SDLV model structure according to an embodiment of this application;

[0044] Figure 3 This is a flowchart illustrating a method for training the model structure to obtain the quality extraction model according to an embodiment of this application.

[0045] Figure 4 This is a schematic diagram of the method for creating a quality extraction model and a soft measurement method according to an embodiment of this application;

[0046] Figure 5 This is a block diagram of a soft measurement device based on a supervised multi-latent variable structure, as described in an embodiment of this application. Detailed Implementation

[0047] This application describes several embodiments, but these descriptions are exemplary and not restrictive, and it will be apparent to those skilled in the art that many more embodiments and implementations are possible within the scope of the embodiments described herein. Although many possible combinations of features are shown in the drawings and discussed in the detailed description, many other combinations of the disclosed features are also possible. Unless specifically limited, any feature or element of any embodiment may be used in combination with, or may replace, any feature or element of any other embodiment.

[0048] This application includes and contemplates combinations of features and elements known to those skilled in the art. The embodiments, features, and elements disclosed in this application may also be combined with any conventional features or elements to form a unique inventive scheme as defined by the claims. Any feature or element of any embodiment may also be combined with features or elements from other inventive schemes to form another unique inventive scheme as defined by the claims. Therefore, it should be understood that any feature shown and / or discussed in this application may be implemented individually or in any suitable combination. Therefore, the embodiments are not limited except by the limitations imposed by the appended claims and their equivalents. Furthermore, various modifications and changes may be made within the scope of the appended claims.

[0049] Furthermore, in describing representative embodiments, the specification may have presented methods and / or processes as a specific sequence of steps. However, the method or process should not be limited to the specific order of steps described herein, to the extent that it does not depend on such a specific order. As will be understood by those skilled in the art, other sequences of steps are also possible. Therefore, the specific order of steps set forth in the specification should not be construed as a limitation of the claims. Moreover, the claims concerning the method and / or process should not be limited to the steps performed in the written order, and those skilled in the art will readily understand that these orders can be varied and still remain within the spirit and scope of the embodiments of this application.

[0050] This application provides a soft measurement method based on a supervised multi-latent variable structure, such as... Figure 1 As shown, the method may include steps S101-S102:

[0051] S101. Input the process data obtained during the thermal power generation scheduling process into the pre-established quality extraction model;

[0052] S102. Extract fuel quality-related information using the quality extraction model; the quality-related information includes process variables and quality variables; wherein, the quality extraction model is a supervised multi-latent variable structure mathematical prediction model (e.g., supervised dual latent variable structure model SDLV), the multi-latent variables include: dynamic latent variables and static latent variables, the dynamic latent variables are used to extract the cross-correlation relationship between the process variables and the quality variables, and the static latent variables are used to retain the cross-correlation relationship between the process variables and the quality variables.

[0053] In an exemplary embodiment of this application, the process data includes, but is not limited to, the oxygen content in flue gas, the carbon content in slag, and the excess oxygen content in the boiler.

[0054] In current technology, there are SFA (Supervised Factor Analysis) models and SLDS (Supervised Linear Dynamics System) models.

[0055] The basic model of SFA is factor analysis (FA). FA is a linear Gaussian latent variable model, which is defined as follows: given latent variable h... t Under the condition of observing variable x t The covariance matrix of the conditional probability distribution is a diagonal matrix, not an isotropic covariance matrix, that is:

[0056] P(x t |h t )=N(x t |Bh t +e x , ∑ x (1)

[0057] Building upon FA, and incorporating information about quality variables, we still use a low-dimensional latent variable h. t The SFA model reflects the main feature information in the original sample dataset. Its mathematical expression is as follows:

[0058] x t =Bh t +e x ,e x ~N(0,∑ x (2)

[0059] y t =Dh t +e y ,e y ~N(0,∑ y (3)

[0060] Equation (2) reflects the conditional probability distribution of equation (1), and equation (3) establishes the unique latent variable h. tWith quality variable y t The relationship between h t Also known as latent variables related to quality, x can be extracted. t and y t The interrelationships between them are applied to the modeling of soft measurement.

[0061] It is worth noting that in SFA h t It is defined as a static latent variable and follows a normal distribution h. t ~N(0,I).

[0062] Compared to SFA, SLDS has the advantage of having only one latent variable h. t Based on this, a constraint related to the time series was added, and a dynamic latent variable q was proposed. t Two latent variables that are sequentially adjacent in time have a first-order Markov relationship, specifically manifested as the relationship between the current state q and the variable q. t , by q t-1 The distribution derivation is as follows, and the relationship between the two is determined by the state transition matrix F. The mathematical expression for SLDS is as follows:

[0063] x t =Aq t +e x ,e x ~N(0,∑ x (4)

[0064] y t =Cq t +e y ,e y ~N(0,∑ y (5)

[0065] q t =Fq t-1 +e q ,e q ~N(0,∑ q (6)

[0066] Due to the existence of a first-order Markov chain, SLDS can extract quality-related dynamic information from process data.

[0067] Cross-correlation, autocorrelation, and uncertainty (measurement noise, missing variables) are the main characteristics of thermal power generation scheduling data, and current technologies mostly address specific problems. However, in most cases, these data characteristics coexist in a single thermal power generation scheduling process, making it difficult for existing soft measurement models to simultaneously account for all of these issues within a single thermal power generation scheduling process.

[0068] For SFA, its latent variable ht Although it can capture some of the process variables x t The cross-correlation between the latent variable q and the quality variable y is considered, but each time slice is independent, ignoring the autocorrelation between latent variables. Therefore, its soft measurement application is limited to static processes. SLDS attempts to simultaneously capture the latent variable q. t With process variable x t and quality variable y t Cross-correlation between them and q t Autocorrelation over time. However, the unique latent variable structure of SLDS is likely to weaken or even lose its original static relationships after the addition of dynamic constraints. Faced with missing available sample data (missing quality variables), existing data downsampling rate modeling methods based on SFA or SLDS [PLVR (downsampling-SFA), DPLVM (downsampling-SLDS)] experience a significant performance drop after data reduction. Experimental results show that downsampling modeling not only discards a large amount of sample data and the process information contained within it, but also severely disrupts the autocorrelation relationships between samples.

[0069] In an exemplary embodiment of this application, to consider the autocorrelation and cross-correlation of variables in an industrial engineering project, and to maintain the performance of the soft measurement model even when a large number of key variables are missing, this embodiment introduces a probabilistic dynamic latent variable (i.e., dynamic latent variable) and a probabilistic static latent variable (i.e., static latent variable). The proposed model SDLV can completely retain quality-related information in two latent feature spaces. That is, when extracting quality-related information, it can simultaneously consider both dynamic and static data features, capturing the cross-correlation relationship between process variables and quality variables from multiple feature directions. Furthermore, measurement noise in the process data is reasonably considered. In addition, considering the lack of sufficient quality variable labels in the actual training samples used for soft measurement modeling, to improve the model's practicality in actual thermal power generation scheduling, the SDLV model is further extended to a semi-supervised form (SSDLV) to adapt to soft measurement modeling with missing quality variable values. Compared to the traditional single latent variable soft measurement model, it divides process information into dynamic and static parts, ensuring the acquisition of quality variable information from multiple directions. The dynamic latent variable is used to extract dynamic information related to quality, while the static latent variable can provide a useful supplement to the retention of quality-related information. Therefore, SDLV's soft measurement performance is absolutely superior in the above models.

[0070] In the exemplary embodiments of this application, the mathematical symbols used in the embodiments of this application will first be explained:

[0071] h t q t These represent the static latent variable vectors respectively. and dynamic latent variable vector s and l are positive integers, representing the number of variables they contain;

[0072] x t y t These represent the process variable vectors that make up the set of observed variables. and mass variable vector These can also be called observation process variables and observation quality variables; m and n are positive integers, representing the number of variables they contain;

[0073] e x e y e q This represents the noise terms. They follow a Gaussian distribution: e x ~N(0,∑ x ), e y ~N(0,∑ y ), e q ~N(0,∑ q ), where ∑ x ,∑ y ,∑ q Let be the covariance matrix of the noise term;

[0074] This represents the state transition matrix, which determines the latent variable q. t The dynamic relationship between them;

[0075] This represents the observation transition matrix, which determines the transformation relationship between latent variables and observations;

[0076] They represent q respectively t and h t The conditional probability distribution (posterior probability distribution); in particular, For q t The initial distribution is 1;

[0077] ω t This is a sampling indicator used to indicate the absence of a quality variable; if the quality variable exists at time t, then ω... t =1, otherwise ω t =0.

[0078] In an exemplary embodiment of this application, creating the quality extraction model may include:

[0079] Create a model structure with multiple latent variables;

[0080] The model structure is trained to obtain the quality extraction model.

[0081] In an exemplary embodiment of this application, creating a model structure with multiple latent variables may include:

[0082] Within the Gaussian framework, the dynamic latent variables are determined by a first-order Markov chain; the static latent variables are independent of each other.

[0083] The dynamic latent variables and the static latent variables are linearly combined to form the process variables and the quality variables.

[0084] In an exemplary embodiment of this application, a novel model structure SDLV with two different types of latent variables is proposed, such as Figure 2 As shown, this enables concurrent modeling of multiple features during the process. The dynamic latent variables of the SDLV model... It can be determined by a first-order Markov chain, while the static latent variables... They are independent of each other. These two latent variables are combined linearly to form process variables. and quality variables Furthermore, the SDLV model is also performed within a Gaussian framework, except for the noise term, h t ~N(0,I), q1 is q t＝1 .

[0085] In an exemplary embodiment of this application, the mathematical expression for SDLV can be represented as follows:

[0086] x t =Aq t +Bh t +e x ,e x ~N(0,∑ x (7)

[0087] y t=Cq t +Dh t +e y ,e y ~N(0,∑ y (8)

[0088] q t =Fq t-1 +e q ,e q ~N(0,∑ q (9)

[0089] In the exemplary embodiments of this application, in the SDLV structure described above, the two types of latent variables in the SDLV simultaneously share common information between process variables and quality variables, and fully reflect the relevant information about quality variables in two different subspaces. That is, the cross-correlation between process variables and quality variables can be captured not only in the dynamic relationship of process data (dynamic latent variables), but also preserved in the static relationship (static latent variables).

[0090] In an exemplary embodiment of this application, the parameter set of the SDLV model described above can be:

[0091]

[0092] During model training, the Expectation Maximization (EM) algorithm is used to maximize the log-likelihood function of the SDLV model to solve for the model parameters.

[0093] In an exemplary embodiment of this application, the model maximizing the log-likelihood function can be expressed as:

[0094]

[0095] In exemplary embodiments of this application, as Figure 3 As shown, training the model structure to obtain the quality extraction model may include steps S201-S203:

[0096] S201. Collect the quality variables that require soft measurement during the thermal power generation scheduling process and the process variables related to the quality variables that require soft measurement as the training dataset for the model structure.

[0097] S202. Based on the training dataset, the Expectation-Maximization (EM) algorithm is used to maximize the log-likelihood function of the model structure, and the parameters of the model structure are solved.

[0098] In an exemplary embodiment of this application, before solving for the parameters of the model structure by maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm based on the training dataset, the method may further include:

[0099] Determine whether there are missing quality variables in the training dataset;

[0100] If the training dataset has missing quality variables, then based on the semi-supervised algorithm, the Expectation Maximization (EM) algorithm is used to maximize the log-likelihood function of the model structure according to the training dataset, and the parameters of the model structure are solved.

[0101] If the training dataset does not have missing quality variables, then based on the fully supervised algorithm, the Expectation-Maximization (EM) algorithm is used to maximize the log-likelihood function of the model structure according to the training dataset, and the parameters of the model structure are solved.

[0102] In exemplary embodiments of this application, as Figure 4 As shown in the diagram, the two branches during the model training phase correspond to the two modeling methods, SDLV and SSDLV, respectively. These two branches will ultimately produce two sets of model parameter Θ suitable for different observation data conditions. SDLV Θ SSDLV If the training dataset has missing quality variables, SSDLV modeling is performed; if the training dataset does not have missing quality variables, SDLV modeling is performed.

[0103] In an exemplary embodiment of this application, the likelihood function of the SSDLV model is modified based on equation (11):

[0104]

[0105] In an exemplary embodiment of this application, the step of maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm based on the training dataset to solve for the parameters of the model structure may include: iteratively executing the following steps until a preset iteration stopping condition is met to obtain the updated parameters as the final parameters of the model structure:

[0106] The posterior probability distributions of the dynamic latent variables and the static latent variables are calculated using the E-step, and the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic latent variables and the static latent variables is derived.

[0107] Based on the posterior probability distribution and its form, the parameter update of the model structure is completed in M ​​steps.

[0108] In an exemplary embodiment of this application, the EM algorithm consists of an E-step and an M-step, which always guarantees that the log-likelihood value of the SDLV model does not decrease while finding the optimal solution for the model parameters. The E-step requires estimating the latent variable q. t and h t We derive the posterior expectation of the SDLV model's log-likelihood function with respect to these two latent variables, and then derive the expected form of the log-likelihood function of the SDLV model with respect to these two latent variables. Given the observed variable x... t＝1：T and y t＝1：T The conditional distribution (posterior probability distribution) of the latent variables can be calculated by the improved Kalman filter algorithm after state prediction and observation correction.

[0109] In an exemplary embodiment of this application, the E-step and M-step will be iterated repeatedly until the log-likelihood function of the model structure satisfies a preset iteration stopping condition (e.g., reaching iteration convergence or reaching a preset number of iterations), and the results of the last calculated E-step and M-step will be used as the true latent variable values ​​and model parameter values ​​of the model.

[0110] In an exemplary embodiment of this application, the step of calculating the posterior probability distributions of the dynamic latent variables and the static latent variables using the E-step, and deriving the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic latent variables and the static latent variables, may include:

[0111] During the state prediction phase, based on the dynamic latent variable q from the previous time t-1... t-1 The posterior probability distribution is used to calculate the dynamic latent variable q at the current time t. t The posterior probability distribution;

[0112] Obtain the dynamic latent variable q at the current time t t and static latent variable h t The joint distribution representation of is used as the form of the posterior probability distribution;

[0113] During the observation correction phase, the dynamic latent variable q is adjusted based on the known process variables and observation quality variables. t and static latent variable h t The joint distribution representation is modified to serve as the modified posterior probability distribution form.

[0114] According to the modified dynamic latent variable q t and static latent variable h t The joint distribution represents the acquisition of the dynamic latent variable q respectively. t and the static latent variable h t The marginal distribution of is used as the posterior probability distribution of the input for the next state prediction stage.

[0115] In an exemplary embodiment of this application, the latent variable q is estimated in the E-step. t and h t We obtain the posterior expectation (posterior probability distribution) of the model and derive the expected form (posterior probability distribution form) of the model's log-likelihood function with respect to these two latent variables, then calculate its maximum likelihood estimate. The specific calculation and operation steps are as follows:

[0116] During the state prediction phase, the dynamic latent variable q t The probability distribution is derived from the previous time step:

[0117]

[0118] In the formula, and They represent q respectively t The predicted mean and covariance values ​​at time t, and For q t-1 Conditional distribution at time t-1.

[0119] Then, q can be defined t and h t The joint distribution at time t is expressed as:

[0120]

[0121] During the observation correction phase, the observation evolution equation is equivalent to:

[0122]

[0123] Given the observed variable value x t and y t Then, combine latent variables The distribution was corrected:

[0124]

[0125]

[0126]

[0127] In the formula, K f These are intermediate variables; based on this, and given the observed variables, the marginal distributions of the two latent variables are further obtained and used as inputs for the next prediction stage. This represents the estimated value of the sum from t to qh. This represents the estimated value of the sum from t to hq.

[0128]

[0129]

[0130] Note that, unlike in the prediction phase, the two latent variable distributions are no longer independent after sharing observational data. Finally, the posterior expectation (posterior probability distribution) of the latent variables is obtained:

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] The expected form of the log-likelihood function with respect to the latent variables (posterior probability distribution form):

[0137]

[0138] In an exemplary embodiment of this application, the step of updating the parameters of the model structure in M ​​steps based on the posterior probability distribution and the form of the posterior probability distribution may include:

[0139] In the M-step, the partial derivatives of the modified posterior probability distribution and its form are calculated, and the partial derivatives are set to zero to calculate the parameters of the model structure.

[0140] In an exemplary embodiment of this application, the model parameters are updated in M ​​steps. The specific calculation and operation steps are as follows to further maximize the completion of the model parameter update:

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151] in, For the updated For the updated F new For the updated F, For the updated ∑ q A new For the updated A and B new For the updated B, C new For the updated C and D new For the updated D, For the updated ∑ x , For the updated ∑ y .

[0152] In an exemplary embodiment of this application, it can be determined whether the log-likelihood function has converged (or reached a certain number of iterations). If this condition is met, the optimal model parameters are determined; otherwise, the next iteration begins, and the posterior expectations of the dynamic and static latent variables are recalculated using the E-step. The expected form of the log-likelihood function of the model structure with respect to the dynamic and static latent variables is derived. Based on the posterior expectation and the expected form, the parameters of the model structure are updated using the M-step. At this point, the E-step utilizes the parameters updated in the previous M-step iteration.

[0153] S203. Substitute the solved parameters into the model structure to obtain the quality extraction model.

[0154] In an exemplary embodiment of this application, the model parameters finally obtained in step S202 are substituted into the established model structure to obtain the quality extraction model, which can be used for SSDLV modeling or SDLV modeling.

[0155] In an exemplary embodiment of this application, the extraction of fuel mass-related information through the mass extraction model may include:

[0156] Calculate the dynamic and static latent variables corresponding to each process data based on the parameters of the quality extraction model;

[0157] The predicted value of the quality variable of the process data is calculated based on the probability distribution of the dynamic and static latent variables corresponding to each process data.

[0158] In an exemplary embodiment of this application, a query sample is introduced. (i.e., the input process data), k refers to time k. Using the determined final model parameters, calculate the two latent variables (dynamic latent variables and static latent variables) corresponding to each query sample. The specific calculation steps may include:

[0159]

[0160]

[0161]

[0162] Here, "^" indicates that the corresponding value is an estimated value. For time k The estimated value, For time k The estimated value, For time k-1 The estimated value, For time k The estimated value, For time k The estimated value, For time k-1 The estimated value.

[0163] In an exemplary embodiment of this application, the predicted value of the quality variable of the query sample is derived from the probability distribution of two latent variables, and the calculation formula may include:

[0164]

[0165] in, This represents the predicted value of the quality variable at time k.

[0166] The exemplary embodiments of this application include at least the following advantages:

[0167] 1. Two latent variables (dynamic latent variables and static latent variables) are introduced. The dynamic latent variables maintain the dynamic relationship, while the static latent variables protect the original static relationship between the variables.

[0168] 2. To address the issue of missing quality variables, this application extends the semi-supervised model (SSDLV model) by locally scaling the original likelihood function. Furthermore, observation of the difference between the locally scaled likelihood function and the original likelihood function reveals that the semi-supervised model's description of process dynamics is highly consistent with that of full-data modeling, thereby ensuring the model performance of this invention when data is missing.

[0169] 3. Compared with traditional measurement and inspection methods, the soft measurement application of the quality extraction model in this application embodiment can achieve real-time measurement of quality variables. As the number of query samples (i.e., process data to be inspected each time) increases, the training dataset of the quality extraction model can be further expanded, thereby realizing continuous updating of the quality extraction model. Furthermore, the subsequent cost of deploying the quality extraction model in industrial measurement is relatively low.

[0170] This application also provides a soft measurement device 1 based on a supervised multi-latent variable structure, such as... Figure 5 As shown, it may include a processor 11 and a computer-readable storage medium 12, wherein the computer-readable storage medium 12 stores instructions that, when executed by the processor 11, implement the soft measurement method.

[0171] In the exemplary embodiments of this application, any of the aforementioned soft measurement methods are applicable to this device embodiment, and will not be described in detail here.

[0172] It will be understood by those skilled in the art that all or some of the steps, systems, or apparatuses disclosed above, and their functional modules / units, can be implemented as software, firmware, hardware, or suitable combinations thereof. In hardware implementations, the division between functional modules / units mentioned above does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all components may be implemented as software executed by a processor, such as a digital signal processor or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit (ASIC). Such software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and can be accessed by a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

Claims

1. A soft measurement method based on a supervised multi-latent variable structure, characterized in that, The method includes: The process data obtained during the thermal power generation scheduling process is input into a pre-established quality extraction model; The quality-related information of fuel is extracted using the quality extraction model; the quality-related information includes process variables and quality variables. The quality extraction model is a supervised mathematical prediction model with a multi-latent variable structure. The multi-latent variables include dynamic latent variables and static latent variables. The dynamic latent variables are used to extract the cross-correlation between the process variables and the quality variables, and the static latent variables are used to retain the cross-correlation between the process variables and the quality variables. The quality extraction model is constructed in the following manner: Creating a model structure with multiple latent variables includes: determining the dynamic latent variables using a first-order Markov chain within a Gaussian framework; the static latent variables being independent of each other; and the process variables and quality variables being formed by linear combination of the dynamic latent variables and the static latent variables. Training the model structure to obtain the quality extraction model includes: collecting quality variables requiring soft measurement during the thermal power generation scheduling process and process variables related to the quality variables requiring soft measurement as training datasets for the model structure; determining whether there are missing quality variables in the training dataset; if there are missing quality variables in the training dataset, then using a semi-supervised algorithm, based on the training dataset, maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm to solve for the parameters of the model structure; if there are no missing quality variables in the training dataset, then using a fully supervised algorithm, based on the training dataset, maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm to solve for the parameters of the model structure; and substituting the solved parameters into the model structure to obtain the quality extraction model.

2. The soft measurement method based on supervised multi-latent variable structure according to claim 1, characterized in that, The step of maximizing the log-likelihood function of the model structure using the Expectation-Maximization (EM) algorithm based on the training dataset to solve for the parameters of the model structure includes: iteratively executing the following steps until a preset iteration stopping condition is met to obtain the updated parameters as the final parameters of the model structure: The posterior probability distributions of the dynamic latent variables and the static latent variables are calculated using the E-step, and the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic latent variables and the static latent variables is derived. Based on the posterior probability distribution and its form, the parameter update of the model structure is completed in M ​​steps.

3. The soft measurement method based on supervised multi-latent variable structure according to claim 2, characterized in that, The step of calculating the posterior probability distributions of the dynamic and static latent variables using the E-step, and deriving the posterior probability distribution form of the log-likelihood function of the model structure with respect to the dynamic and static latent variables, includes: During the state prediction phase, based on the dynamic latent variables of the previous time t-1... The posterior probability distribution is used to calculate the dynamic latent variables at the current time t. The posterior probability distribution; Obtain the dynamic latent variables at the current time t and static latent variables The joint distribution representation of is used as the form of the posterior probability distribution; During the observation correction phase, the dynamic latent variables are adjusted based on the known observation process variables and observation quality variables. and static latent variables The joint distribution representation is modified to serve as the modified posterior probability distribution form. According to the revised dynamic latent variables and static latent variables The joint distribution represents the acquisition of the dynamic latent variables respectively. and the static latent variables The marginal distribution of is used as the posterior probability distribution of the input for the next state prediction stage.

4. The soft measurement method based on supervised multi-latent variable structure according to claim 2, characterized in that, The parameter update of the model structure based on the posterior probability distribution and its form, using M steps, includes: In the M-step, the partial derivatives of the modified posterior probability distribution and its form are calculated, and the partial derivatives are set to zero to calculate the parameters of the model structure.

5. The soft measurement method based on supervised multi-latent variable structure according to claim 1, characterized in that, The extraction of fuel quality-related information through the quality extraction model includes: Calculate the dynamic and static latent variables corresponding to each process data based on the parameters of the quality extraction model; The predicted value of the quality variable of the process data is calculated based on the posterior probability distribution of the dynamic and static latent variables corresponding to each process data.

6. A soft measurement device based on a supervised multi-latent variable structure, comprising a processor and a computer-readable storage medium, wherein the computer-readable storage medium stores instructions, characterized in that, When the instruction is executed by the processor, the soft measurement method based on a supervised multi-potential variable structure as described in any one of claims 1-5 is implemented.

Images