CN111650894A - Bayesian network complex industrial process soft measurement method based on hidden variables - Google Patents

Abstract

The invention discloses a Bayesian network complex industrial process soft measurement method based on hidden variables. The method gives full play to the advantages of the Bayesian network and local weighted learning, weights the original data by calculating the global similarity of the online sample to be predicted and the corresponding training sample through the hidden variables, inputs the weighted data into the Bayesian network for prediction, and realizes the self-adaptive soft measurement of the complex non-stationary industrial process. Aiming at the time-varying characteristic of the complex industrial process, the similarity calculation and weighting based on the hidden variable are introduced on the basis of supervised training of the Bayesian network, so that the model overfitting phenomenon is relieved, the prediction precision is improved, and method support is provided for soft measurement modeling of the quality variable closely related to the production safety, the production quality and the production efficiency in the complex chemical production process.

Description

Bayesian network complex industrial process soft measurement method based on hidden variables

Technical Field

The invention belongs to the field of continuous chemical process control and soft measurement, and particularly relates to a Bayesian network complex industrial process soft measurement method based on hidden variables.

Background

In continuous chemical industrial production, production states exist which are difficult to measure and monitor on line, and in the face of monitoring of relevant variables in such complex processes, soft measurement methods are often used: the sample data of the physical quantity which is easy to detect is sampled first, and then the variables which can reflect the state of the production process and the quality of the product to some extent are indirectly estimated. At present, various models based on data driving are applied to soft measurement, and a hidden variable which can reveal the close relation between a quality variable and a production system is difficult to be effectively found out by a learning method directly based on the data driving.

The Bayesian network is one of effective theoretical models in the field of uncertain knowledge and reasoning at present, the graph model combining the graph theory and the probability theory can better process complex, fuzzy and uncertain scenes, and the graph model can be used for constructing a soft measurement model to be a research hotspot.

In the actual complex industrial process, the model is degraded due to the aging of platform equipment, the inactivation of the catalyst, the change of the process environment and the like, and the model established originally is not suitable for the existing operation state any more, so that the precision of the model is reduced. In order to correctly track the process state and solve the time-varying problem, a bayesian network based on a sliding window and instant learning is proposed, but certain limitations exist, and the influence of hidden variables on a soft measurement model is often ignored. Aiming at the time-varying process in the complex industrial production, the similarity calculation and weighting based on the hidden variable are introduced on the basis of supervised Bayesian network training, so that the overfitting is relieved, the model error is reduced, and the prediction accuracy of the quality variable in the complex industrial process is improved.

Disclosure of Invention

In order to solve the technical problems in the background technology, the invention provides a Bayesian network complex industrial process soft measurement method based on hidden variables.

The technical scheme adopted by the invention is as follows:

step 1, collecting data closely related to production safety and quality in a complex continuous industrial process through a sensor, and dividing the data into a training sample set [ X, Y ]]And query sample set [ X ]_query,Y_query]. Training sample set, process variable X ═ X₁,x₂,...x_n]∈R^m×nAnd the mass variable Y ═ Y₁,y₂,...y_n]∈R^k×nWill be used for soft measurement modeling, where m is a process variationThe number of variables of the quantity X, k the number of variables of the quality variable Y, and n the number of samples included in the data set. Query sample set, Process variable X_queryFor soft measurement prediction, Y_queryFor verifying the soft measurement prediction effect;

the data related to the complex industrial process may specifically include values of important variables collected by a sensor installed in a chemical reaction vessel of the complex industrial process. Process variables are variables that are relatively easy to monitor, and quality variables are important variables in relation to the quality and safety of the production. The sensors include temperature sensors mounted at the top, the column plate and the bottom of the column, pressure sensor at the overhead gas, velocity sensor at the reflux drum and velocity sensor at the next stage of production inlet.

Step 2, for each query sample x_q∈X_querySelecting a certain number of samples D from the training sample set by using a sliding window_trainAccording to the relative variable relationship of the industrial process and D_trainThe size of the system is constructed with a supervision Bayesian network, the Bayesian network is trained, and D is obtained_trainCorresponding hidden variable T, x_qInput Bayes network to obtain hidden variable t_new；

Step 3, utilizing the hidden variable T, T_newFinding the confidence and local similarity of each sliding window to obtain x_qAnd X_trainThe global similarity of (a) is used to locally weight the original industrial process variable;

step 4, taking the global similarity as a weight pair X_train、Y_trainWeighting to obtain training data X after local weighting_t、Y_tIs mixing X_t、Y_tInputting Bayes network according to node to obtain x_qCorresponding predicted value y_qFinish all the query samples x_qA prediction of a corresponding industrial process quality variable;

and 5, after all the query samples are predicted, measuring the prediction effect of the soft measurement model on the complex industrial process variables. The accuracy of the prediction result is measured by the root mean square error RMSE, and R is used²And measuring the data tracking capability of the prediction result, wherein the calculation method comprises the following steps:

wherein y is_realI.e. each query sample x_qCorresponding to y_q，

Is y_realHas an average value of y_predAnd n is the number of the query samples.

The step 2 specifically comprises the following steps:

step 2-1, traversing the training sample set divided in step 1 by using a sliding window, wherein the number of samples included in each window is W, the number of windows is s, and the number of traversed samples is W-W × s, that is, each query sample x is_q∈X_queryThe corresponding training set is D_train＝[X_train,Y_train]∈R^(m+k)×W；

Step 2-2, selecting sequence α ═<X,t,Y>Constructing a quality-related Bayesian network with a node number of 3, which comprises a quality variable Y in a training set, relative to a conventional Bayesian network_trainAnd (4) supervision. Inputting X according to corresponding network node_train、Y_trainI.e. supervised training of the Bayesian network and solving of the training data set D_trainCorresponding hidden variable T ∈ R^W×iAnd i represents the number of hidden variables;

the Bayesian network of the present invention is schematically illustrated in FIG. 1, wherein nodes represent variables, and directional arrows between nodes represent causal dependencies between variables, the Bayesian network having 3 nodes in total, three directional edges, and an order of α<X,t,Y>Node X represents a process variable, node T represents a hidden variable, and node Y represents a quality variable. The node X is a father node, two edges start to respectively point to child nodes T and Y, and one edge of the node T starts to point to Y. Network node X corresponds to input X_trainNode Y corresponds to input Y_trainTraining BayesAnd after the network is started, the hidden variable is output from the node T.

Step 2-3, new process variable x_qInputting a Bayesian network node X, and calculating X_qCorresponding hidden variable t_new∈R^{1 ×i}。

In the step 3, the flow is as shown in fig. 3, and the specific process is as follows:

step 3-1, applying SVDD to determine x_qThe confidence with each window, the window confidence for the vth is defined as:

wherein a is_vAnd R_vRespectively representing hidden variables T calculated from the v-th window_vCentre coordinates and radius, x, of the constructed hypersphere_vRepresents t_newCorresponding coordinates in the hypersphere.

Step 3-2, supposing that the r training sample belongs to the v moving window, t_newCorresponding to the current training sample to obtain an implicit variable t_rLocal similarity of S_vThe calculation is as follows:

wherein the content of the first and second substances,

to take a tuning parameter of 0-1, σ_dIs d_τThe corresponding standard deviation.

Step 3-3, the r training sample and x in the v window_qThe global similarity of (a) is:

Sim_r＝win_v·S_v

x_qthe global similarity to the training samples can be expressed as:

SIM＝[Sim₁,Sim₂,···,Sim_r,···Sim_w×s]

in the step 4, the specific process is as follows:

step 4-1, calculating X according to the formula shown in step 3_t、Y_t：

Step 4-2, inputting X according to the corresponding network node_t、Y_tTraining Bayesian network to find x_qCorresponding predicted value y_q；

Step 4-3, repeating the step 2 to the step 4 until the prediction is finished X_queryPredicted value Y corresponding to all the query samples in the database_qAnd step 5 is executed.

Compared with the prior art, the invention has the following beneficial effects:

1. the causal dependence relationship among the variables is fully excavated by utilizing the process knowledge of industrial production, and the Bayesian network is constructed, so that the relationship among the variables is more visual.

2. The Bayesian network is supervised and trained by fully utilizing process knowledge and important labeled production data.

3. Latent relations among a plurality of variables are mined by calculating latent variables, the latent variables are used for weighting original data, the utilization rate of the original data is improved, model updating is facilitated, and adaptive soft measurement of an industrial process is realized

Aiming at the time-varying characteristic of the actual production process, the Bayesian network is used for calculating the hidden variable in a supervision manner, the traditional Bayesian network is expanded into a self-adaptive soft measurement model, and fewer training samples can be selected for prediction; compared with other traditional self-adaptive soft measurement models, the method has the advantages that overfitting is relieved, prediction accuracy is improved, and method support is provided for soft measurement modeling of quality variables closely related to production safety, production quality and production efficiency in a complex chemical production process.

Drawings

FIG. 1 is a Bayesian network architecture to which the present invention relates;

FIG. 2 is a schematic view of a debutanizer configuration in accordance with the present invention;

FIG. 3 is a schematic flow chart of the method of the present invention;

FIG. 4 is a graph showing the results of predicting butane content based on the method of the present invention;

FIG. 5 is a graph showing the results of a comparative method for predicting butane content using a hidden variable-free weighted Bayesian network;

FIG. 6 is a process scheme according to the present invention.

Detailed Description

The invention is further illustrated by the following figures and examples.

The embodiment of the invention and the implementation process thereof are as follows:

taking the production process of the debutanizer as an example, the soft measurement modeling method is described in detail based on the data recorded in the operation process, and the technical route is shown in fig. 6.

The debutanizer is a refining process for naphtha cracking, and aims to separate butane from naphtha and verify a reference platform of different algorithms in the field of soft measurement models. FIG. 2 is a process flow diagram for debutanizer production. In the debutanizer column, the detection of the butane content at the bottom of the column is a vital part in monitoring the production. However, butane is not as easy to measure as temperature. The present invention therefore selects readily measurable process variables related to temperature, pressure, flow, concentration as shown in table 1 to estimate the mass variable, i.e., butane content.

TABLE 1

Process variable	Description of variables
		U1	Temperature at the top of the column
U2	Pressure at the top of the column
		U3	Amount of reflux
U4	Next stage flow
		U5	Temperature of column plate 6
U6	Temperature at the bottom of the tower (left)
		U7	Temperature at the bottom of the tower (Right)

As shown in fig. 3, in order to predict the butane content at the bottom of the debutanizer tower by applying a complex industrial process soft measurement modeling method based on supervised bayesian network hidden variable similarity and local weighting, the following steps are made:

step 1, collecting variables closely related to production safety and quality in a complex continuous industrial process through a sensor, and dividing the variables into a training sample set [ X, Y ]]And query sample set [ X ]_query,Y_query]. In the sample set, the number of process variables is 7, the number of quality variables is 1, and the number of query samples is 500.

Step 2, for each query sample x_q∈X_querySelecting 50 training samples by using 10 sliding windows with the window size of 5, inputting the training samples into a Bayesian network, performing parameter learning by using an EM (effective man algorithm), and then solving a hidden variable T corresponding to the training samples by using a combined tree inference engine; will be asx_qInputting evidence into network to obtain hidden variable t_new；

Step 3, utilizing the hidden variable T, T_newFinding the confidence and local similarity of each sliding window to obtain x_qAnd X_trainThe global similarity of (a) is used to locally weight the original industrial process variable;

step 4, taking the global similarity as a weight pair X_train、Y_trainWeighting to obtain training data X after local weighting_t、Y_tIs mixing X_t、Y_tInputting Bayes network according to node to obtain x_qCorresponding predicted value y_qFinish all the query samples x_qA prediction of a corresponding industrial process quality variable;

and 5, after all the query samples are predicted, measuring the prediction effect of the soft measurement model on the complex industrial process variables. The accuracy of the prediction result is measured by the root mean square error RMSE, and R is used²And measuring the data tracking capability of the prediction result, wherein the calculation method comprises the following steps:

wherein y is_realI.e. each query sample x_qCorresponding to y_q，

Is y_realHas an average value of y_predAnd n is the number of the query samples.

The result of the method of the invention for predicting butane content is shown in fig. 4, and the traditional method comprises the following steps: local weighted Bayes soft measurement modeling without introducing hidden variables is used as an effect comparison method, the prediction results are shown in FIG. 5, the prediction effect pairs of the two methods are shown in Table 2, and as can be seen from Table 2, the method of the invention has higher prediction accuracy than a Bayes network without introducing hidden variables.

TABLE 2

	The method of the invention	Conventional methods
			RMSE	0.0350	0.0536
R2	0.9414	0.8622

In summary, the method provided by the invention is a Bayesian network complex industrial process soft measurement method based on hidden variables, which can complete soft measurement modeling of a complex industrial process, realize prediction of quality variables such as butane and improve the prediction accuracy of soft measurement to a certain extent.

Claims (4)

1. A Bayesian network complex industrial process soft measurement method based on hidden variables is characterized by comprising the following steps:

step 1, collecting variables closely related to production safety and quality in a complex continuous industrial process through a sensor, and dividing the variables into a training sample set [ X, Y ]]And query sample set [ X ]_query,Y_query]. Training sample set, process variable X ═ X₁,x₂,...x_n]∈R^{m ×n}And the mass variable Y ═ Y₁,y₂,...y_n]∈R^k×nIt will be used for soft measurement modeling, where m is the number of variables for the process variable X, k is the number of variables for the quality variable Y, and n is the number of samples contained in the data set. Query sample set, Process variable X_queryFor soft measurement prediction of on-line samples to be predicted, Y_queryFor verifying the soft measurement prediction effect;

step 2, for each query sample x_q∈X_querySelecting a certain number of samples D from a training sample set by using a sliding window_trainAccording to the relative variable relationship of the industrial process and D_trainThe size of the system is constructed with a supervision Bayesian network, the Bayesian network is trained, and D is obtained_trainCorresponding hidden variable T, x_qInput Bayes network to obtain hidden variable t_new；

Step 3, utilizing an implicit variable T, t_newThe confidence coefficient and the local similarity of each sliding window are calculated, and then x is obtained_qAnd X_trainThe global similarity of (a) is used to locally weight the original industrial process variable;

step 4, taking the global similarity as a weight pair X_train、Y_trainWeighting to obtain training data X after local weighting_t、Y_tIs mixing X_t、Y_tInputting Bayes network according to node to obtain x_qCorresponding predicted value y_qFinish all the query samples x_qA prediction of a corresponding industrial process quality variable;

and 5, after all the query samples are predicted, measuring the prediction effect of the soft measurement model on the complex industrial process variables. The accuracy of the prediction result is measured by the root mean square error RMSE, and R is used²And measuring the data tracking capability of the prediction result, wherein the calculation method comprises the following steps:

wherein y is_realI.e. each query sample x_qCorresponding to y_q，

Is y_realHas an average value of y_predAnd n is the number of the query samples.

2. The Bayesian network complex industrial process soft measurement method based on implicit variables as recited in claim 1, wherein: the step 2 specifically comprises:

step 2-1, traversing the training sample set divided in step 1 by using a sliding window, wherein the number of samples included in each window is W, the number of windows is s, and the number of traversed samples is W-W × s, that is, each query sample x is_q∈X_queryThe corresponding training set is D_train＝[X_train,Y_train]∈R^(m+k)×W；

Step 2-2, selecting sequence α ═<X,t,Y>Constructing a quality-related Bayesian network with a node number of 3, which comprises a quality variable Y in a training set, relative to a conventional Bayesian network_trainAnd (4) supervision. Inputting X according to corresponding network node_train、Y_trainI.e. supervised training of the Bayesian network and solving of the training data set D_trainCorresponding hidden variable T ∈ R^W×iAnd i represents the number of hidden variables;

step 2-3, new process variable x_qInputting into Bayesian network to obtain x_qCorresponding hidden variable t_new∈R^1×i。

3. The Bayesian network complex industrial process soft measurement method based on implicit variables as recited in claim 1, wherein: in the step 3, the specific process is as follows:

step 3-1, applying SVDD to determine x_qThe confidence with each window, the confidence of the vth window, is calculated as follows:

wherein a is_vAnd R_vRespectively representing hidden variables T calculated from the v-th window_vAt the centre coordinate and radius, x, of the constructed hypersphere_vRepresents t_newCorresponding coordinates in the hypersphere.

Step 3-2, supposing that the r training sample belongs to the v moving window, t_newCorresponding hidden variable t to current training sample_rLocal similarity of S_vThe calculation is as follows:

wherein the content of the first and second substances,

to take a tuning parameter of 0-1, σ_dIs d_τThe corresponding standard deviation.

Step 3-3, the r training sample and x in the v window_qThe global similarity of (a) is:

Sim_r＝win_v·S_v

x_qthe global similarity to the training samples can be expressed as:

SIM＝[Sim₁,Sim₂,···,Sim_r,···Sim_w×s]

4. the Bayesian network complex industrial process soft measurement method based on implicit variables as recited in claim 1, wherein: in the step 4, the specific process is as follows:

step 4-1, calculating X according to the formula shown in step 3_t、Y_t：

In the step 4-2, the step of the method,inputting X according to corresponding network node_t、Y_tTraining Bayesian network to find x_qCorresponding predicted value y_q；

Step 4-3, repeating the step 2 to the step 4 until the prediction is finished X_queryPredicted value Y corresponding to all the query samples in the database_qAnd step 5 is executed.

Images