CN102880151A - Double-layer data model-driven plant-level chemical process monitoring method - Google Patents

Abstract

The invention discloses a double-layer data model-driven plant-level chemical process monitoring method. Relationships between sub-modules are modeled by constructing a double-layer data statistical analysis model on the basis of blocking modeling, so that a plant-level process is globally monitored. Compared with the conventional other plant-level process monitoring methods, the method has the advantages that each unit of the process can be monitored in each sub-module, relationship information between the sub-modules of the plant-level process can be effectively combined, and the whole plant-level process is globally monitored by utilizing a second-layer data model, so that the monitoring performance of the plant-level chemical process is greatly improved, and the industrial automation of the plant-level process can be favorably expanded and implemented.

Description

The level of factory Monitoring of Chemical method that double-deck data model drives

Technical field

The invention belongs to safety monitoring and the field of quality control of chemical engineering industry production run, particularly a kind of level of factory Monitoring of Chemical method of double-deck data model driving.

Background technology

Along with the production-scale expanding day of modern chemical industry process, a kind of novel level of factory process more and more causes people's concern.Also obtained the extensive concern of academia and industry member for modeling, monitoring and the control of level of factory process thereupon.A typical level of factory process comprises a plurality of operating units, a plurality of flow path device usually, even includes a plurality of operations workshop.Like this complicated large-size chemical process adopts traditional monitoring method based on mechanism model will be very difficult.By contrast, owing to having accumulated a large amount of data in the process, the modeling of based on data and monitoring method are fit to and complex large-sized level of factory process more.Level of factory process monitoring method commonly used all is to be based upon on the basis of data-driven model mostly at present, in these large class methods, usually first whole level of factory process is divided into a plurality of submodules according to operating unit and function, the monitoring model of then setting up based on data for different submodules respectively carries out the piecemeal monitoring to process.But these class methods have often been ignored the mutual relationship between the submodule, and this relation might have influence on the Monitoring Performance of whole level of factory process.Therefore, if can on the basis of submodule data modeling, set up a model to the mutual relationship between them again, will greatly improve data model to the capturing ability of level of factory procedural information.Simultaneously, based on the monitoring system of a double-deck data model, will make simultaneously Real-Time Monitoring to the overall situation and the local mode of operation of level of factory process, the robotization that is very beneficial for the level of factory process is implemented.

Summary of the invention

The object of the invention is to the deficiency for existing level of factory process monitoring method, the level of factory Monitoring of Chemical method that provides a kind of double-deck data model to drive.

The objective of the invention is to be achieved through the following technical solutions: the level of factory Monitoring of Chemical method that a kind of double-deck data model drives may further comprise the steps:

(1) data of utilizing Distributed Control System (DCS) to collect the level of factory chemical process form the training data sample set that modeling is used: X ∈ R ^{N * m}Wherein, n is the number of sample data collection, and m is the variable number of sample data collection.

(2) with whole level of factory process data collection X ∈ R ^{N * m}Be divided into different submodules

B=1 wherein, 2 ..., B is the sequence number of submodule, B is total submodule number.Deposit these submodule data sets in database backup.

(3) for data set corresponding to each submodule, it is carried out normalized, set up the data statistics monitoring model based on pivot analysis, extract pivot information, set up monitoring and statistics amount and the monitoring limit thereof of submodule.

(4) obtain after the pivot information of each submodule it being re-started arrangement, be combined into new data set and be used for making up second layer data statistic analysis model.

(5) for the new data matrix that mixes pivot information, adopt one-class support vector machines that it is carried out modeling, set up the relational model between each submodule, finish whole modeling and the monitoring of level of factory process.

(6) collect new process data, and it is carried out pre-service and normalization.

(7) new process data is divided into a plurality of submodules, is input in the submodule principal component model, calculate corresponding pivot composition.

(8) reconfigure the pivot information of different submodules, new data matrix is input in the one-class support vector machines model, calculate corresponding monitoring and statistics value, judge the running status of active procedure.

The invention has the beneficial effects as follows: the present invention is by making up double-deck data statistic analysis model to the level of factory process, overcome the defective of traditional piecemeal modeling method, set up the relational model between each submodule in the level of factory process, and introduced new distance statistics amount of one-class support vector machines structure for process monitoring.Compare other present level of factory process monitoring method, the present invention not only can monitor by the unit to process in each submodule, and can be effectively in conjunction with the relation information between each submodule of level of factory process, utilize second layer data model that the overall situation of whole level of factory process is monitored, greatly improve the Monitoring Performance of level of factory chemical process, be very beneficial for expansion and the enforcement of level of factory process industrial robotization.

Description of drawings

Fig. 1 is the process flow diagram of level of factory chemical process;

Fig. 2 is to the monitoring result of normal data set in the level of factory process in the inventive method;

Fig. 3 is to the monitoring result of abnormal data set in the level of factory process in the inventive method;

Fig. 4 is to the monitoring result of abnormal data set in the level of factory process in traditional Principal Component Analysis Model.

Embodiment

The present invention is directed to the monitoring problem of level of factory chemical process, at first utilize Distributed Control System (DCS) to collect the data of this process, it is carried out necessary pre-service and normalization, then whole process data collection is divided into different submodules.For data set corresponding to each submodule, set up respectively one Principal Component Analysis Model, and set up the control limit of monitoring and statistics amount.Obtain after pivot information corresponding to each submodule data set it being reassembled into new data set, then adopt one-class support vector machines that it is carried out modeling.When new process data is monitored, equally these data are divided into different submodules, utilize average and the standard deviation of each submodule modeling data that it is carried out normalized, obtain after the normal data, the Principal Component Analysis Model that recycles each submodule calculates the pivot information of new data in different submodules.All pivot information that obtain in the submodule are reconfigured, be input in the one-class support vector machines model of the second layer, calculate the value apart from the monitoring and statistics amount, realize the overall situation monitoring to the level of factory process.

The key step of the technical solution used in the present invention is as follows respectively:

The data that the first step utilizes Distributed Control System (DCS) to collect the level of factory chemical process form the two-dimentional training sample set that modeling is used: X ∈ R ^{N * m}Wherein, n is the number of sample data collection, and m is the variable number of sample data collection, and R is real number.

Second step is according to the operating unit type and relevant functional attributes of process, the two-dimentional training sample set X ∈ R that step 1 is obtained ^{N * m}Be divided into different submodules:

B=1 wherein, 2 ..., B is the sequence number of submodule, B is total submodule number.Deposit these submodule data sets in database backup.

The 3rd step was carried out normalized for data set corresponding to each submodule to it, set up the data statistics monitoring model based on pivot analysis, extracted pivot information, set up monitoring and statistics amount and the monitoring limit thereof of submodule.

The data set corresponding to each submodule

B=1,2 ..., B carries out pre-service and normalization, and namely so that the average of each process variable is zero, variance is 1, obtains new data matrix collection.The fundamental purpose of this step is for so that the yardstick of process data can not have influence on the result of monitoring.After normalization, the data of various process variable just are under the identical yardstick, can not have influence on follow-up monitoring effect afterwards.Then, this data set is carried out pivot analysis, former space is divided into principal component space and residual error space, choose suitable pivot number, the loading matrix P that obtains _bWith score matrix T _b, and the residual error of acquisition modeling, that is:

$X_b=T_b\·P_b^T+{\tilde{T}}_b\·{\tilde{P}}_b^T$

On the basis of Principal Component Analysis Model, by constructing following two statistic T ²Realize monitoring to each submodule inside with SPE:

$T_b^2=t_b^T{\mathrm{\Λ}}_b{t}_b;$ ${\mathrm{SPE}}_b=e_b^T{e}_b,$

Wherein, t _bBe submodule score matrix T _bIn vector, corresponding to the pivot variable of a sub-module data,

Estimation residual error for these data.For the state to data is monitored, we need to set up

And SPE _bThe monitoring and statistics limit of statistic

And SPE _{B, lim}, that is:

$T_{b,\lim }^2=\frac{k_b(n-1)}{n-k_b}{F}_{k_b,(n-k_b),\mathrm{\α}};$ ${\mathrm{SPE}}_{b,\lim }=g_b{\mathrm{\χ}}_{n_b,\mathrm{\α}}^2;$

Wherein, F represents the statistical distribution of F form, x ²Be x ²The statistical distribution of form, α is statistical confidence, g _b=var (SPE _b)/[2mean (SPE _b)], h _b=2[mean (SPE _b)] ²/ var (SPE _b), mean (SPE wherein _b) and var (SPE _b) be respectively SPE _bThe average of statistic and variance.Respectively to b=1,2 ..., the different submodules of B carry out modeling, finish the ground floor construction of data model.

The 4th step obtained the pivot information of each submodule

B=1,2 ..., after the B, it is re-started arrangement, be combined into new data set and be used for making up second layer data statistic analysis model, that is:

T _com=[T ₁T ₂…T _b…T _B]。

The 5th step adopted one-class support vector machines that it is carried out modeling for the new data matrix that mixes pivot information, set up the relational model between each submodule, finished whole modeling and the monitoring of level of factory process.

Mix pivot data matrix T for two dimension _Com, set up the one-class support vector machines Data Analysis Model.At first utilize nonlinear function that process data is projected in the high-dimensional feature space, namely

One-class support vector machines is set up model by finding the solution following optimal problem:

$\underset{R,a,\mathrm{\ξ}}{\min }{R}^2+C\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\ξ}}_i$

$s.t.{\left|\right|\mathrm{\Φ}\left(t_i\right)-a\left|\right|}^2\≤R^2+{\mathrm{\ξ}}_i,{\mathrm{\ξ}}_i\≥0,i=\mathrm{1,2},\·\·\·,n$

Wherein, R and a are respectively radius and the centre ofs sphere of suprasphere in the high-dimensional feature space, Φ (t _i) be the non-linear projection function, C is the one-class support vector machines parameter, and by this parameter, one-class support vector machines can average out between the mistake minute rate of the volume of suprasphere and sample, and ξ is the slack variable of each sample.In the solution procedure of reality, usually adopt following dual propositions to construct one-class support vector machines, that is:

$\underset{{\mathrm{\α}}_i}{\min }\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_{i}K(t_i,t_j)-\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i{\mathrm{\α}}_{j}K(t_i,t_j);$

$s.t.0\≤{\mathrm{\α}}_i\≤C,\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i=1$

Wherein, K (t _i, t _j)=＜Φ (t _i), Φ (t _j) be kernel function, and usually being chosen for the form of gaussian kernel function, α is coefficient corresponding to each sample, i.e. Lagrange multiplier.The modeling result of one-class support vector machines is: the α value that most of sample is corresponding is zero, only has α value corresponding to the crucial sample of fraction non-vanishing, and these samples are called as support vector.

In high-dimensional feature space, ask for suprasphere the centre of sphere and radius as follows:

$a=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\α}}_i\mathrm{\Φ}\left(t_i\right),$

$R=\sqrt{1-2\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K({\mathrm{<figure-callout\; id="0"\; label="t"\; filenames="HDA00002243042700022.png"\; state="\{\{state\}\}">t</figure-callout>}}_0,t_i)+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_{j}K(t_i,t_j)},$

Wherein, t ₀Be one of them support vector in single class vector machine model.

The 6th step was collected new process data, and it is carried out pre-service and normalization.

The 7th step was divided into a plurality of submodules with new process data, was input in the submodule principal component model, calculated corresponding pivot composition.

At first, with new process data x ^NewBe divided into a plurality of submodules, namely

$x^{\mathrm{new}}=\left[\begin{array}{cccccc}{x}_1^{\mathrm{new}}& x_2^{\mathrm{new}}& \&CenterDot;\&CenterDot;\&CenterDot;& x_b^{\mathrm{new}}& \&CenterDot;\&CenterDot;\&CenterDot;& x_B^{\mathrm{new}}\end{array}\right]$

Different submodule data vectors is input in the corresponding submodule principal component model, calculates corresponding pivot composition as follows

$t_1^{\mathrm{new}}=P_1^T{x}_1^{\mathrm{new}},$ $t_2^{\mathrm{new}}=P_2^T{x}_{12}^{\mathrm{new}},$ ...， $t_B^{\mathrm{new}}=P_B^T{x}_B^{\mathrm{new}}$

The 8th step reconfigured the pivot information of different submodules, and new data matrix is input in the one-class support vector machines model, calculated corresponding monitoring and statistics value, judged the running status of active procedure.

The pivot information combination that each submodule data vector is corresponding is as follows:

$t_{\mathrm{com}}^{\mathrm{new}}=\left[\begin{array}{cccccc}{P}_1^T{x}_1^{\mathrm{new}}& P_2^T{x}_2^{\mathrm{new}}& \&CenterDot;\&CenterDot;\&CenterDot;& P_b^T{x}_b^{\mathrm{new}}& \&CenterDot;\&CenterDot;\&CenterDot;& P_B^T{x}_B^{\mathrm{new}}\end{array}\right],$

On this basis, with the new pivot blended data vector that obtains

Be input in the one-class support vector machines model, in higher dimensional space, calculate this data vector to the distance of the suprasphere centre of sphere, that is:

$D_{\mathrm{new}}=d\left(\mathrm{\&Phi;}\left(t_{\mathrm{com}}^{\mathrm{new}}\right)\right)=\left|\right|\mathrm{\&Phi;}\left(t_{\mathrm{com}}^{\mathrm{new}}\right)-a\left|\right|=\sqrt{1-2\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{i}K(t_{\mathrm{com}}^{\mathrm{new}},t_i)+\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_i{\mathrm{\&alpha;}}_{j}K(t_i,t_j)},$

Work as D _New≤ D _LimDuring=R, illustrate that the running status of level of factory process is good, current do not have fault to occur, on the contrary, if D _NewD _LimDuring=R, illustrate that this monitoring and statistics amount has surmounted its statistics control limit, certain fault might occur in the level of factory process, need to provide warning message to process.

The validity of the inventive method is described below in conjunction with a concrete level of factory chemical process example.The process flow diagram of this process as shown in Figure 1, whole process is comprised of 5 different operating units.Choose the monitoring that 33 main process variablees are used for this level of factory process, as shown in table 1, corresponding informance according to process operation Elementary Function and variable, monitored parameters is divided into 3 different submodules, this is because wherein have the variable of two operating units very few, and and adjacent operating unit coupling is stronger, therefore they is put under in other the operating unit.Gather altogether during the course 960 data points and be used for modeling, next in conjunction with this detailed process implementation step of the present invention is at length set forth:

1. collection process data, the data pre-service, level of factory process piecemeal, the submodule Principal Component Analysis Model makes up (ground floor data model structure)

960 effective normal data samples collecting are carried out pre-service, remove average and variance, and it is divided into three different submodules, respectively corresponding 16,10 and 7 process variable.Like this, the submodule data set that obtains is respectively X ₁∈ R ^{960 * 16}, X ₂∈ R ^{960 * 10}And X ₃∈ R ^{960 * 7}Then, respectively for each submodule data set, set up its Principal Component Statistics analytical model, utilize the cumulative proportion in ANOVA principle, the cumulative data explanation degree that is pivot surpasses 85%, the pivot number of choosing is respectively 11,5 and 3, and cumulative proportion in ANOVA is respectively 88.72%, 91.64% and 90.54%.In addition, the T that obtains of each Principal Component Analysis Model ²With SPE statistics quantitative statistics limit be respectively 25.1908 and first submodule of 8.3297(), 15.2452 and second submodule of 4.9734(), 11.4300 and the 3rd submodule of 5.6033().

2. based on the level of factory process global modeling (second layer data model structure) of one-class support vector machines

The pivot information that each submodule is obtained re-starts combination, obtains following new pivot data matrix, utilizes one-class support vector machines that it is carried out modeling, the position of the centre of sphere of definite suprasphere and the size of radius in higher dimensional space.In the time of Selecting All Parameters, so that a wrong minute rate is controlled at about 1%, the monitoring and statistics amount of gained has just represented 99% confidence limit like this.In high-dimensional feature space, the radius size of the suprasphere that finally obtains is 0.7059.

3. obtain current Monitoring Data information, and it is carried out pre-service and normalization

In order to test the validity of new method, respectively normal data set and abnormal data set are tested, wherein, normal data set comprises 500 data points, abnormal data set comprises 960 data points, and (wherein 160 of fronts are normal data, 800 is abnormal data afterwards), at first be to utilize average and variance in the model bank that it is carried out normalized, so that this test data has identical yardstick with modeling data.

4. the on-line monitoring of level of factory chemical process

Each test data of at first normal data being concentrated is divided into three different submodule data vectors, utilizes Principal Component Analysis Model corresponding to each submodule that it is carried out the pivot information extraction.Obtain after the pivot information of each submodule it being re-started combination, be input in the one-class support vector machines model, in high-dimensional feature space, calculate the distance of itself and the suprasphere centre of sphere, i.e. the value of distance statistics amount, and compare with radius, obtain the Real-Time Monitoring result, as shown in Figure 2.Can be found out by this figure, double-deck data model is to the monitoring result of normal data or satisfied, and the data volume of wrong report about 1%, meets the hypothesis of modeling substantially.Next abnormal data set is monitored, the result as shown in Figure 3, as can be seen from Figure, after 160 sampled points, double-deck data model has successfully monitored the unusual of level of factory process, and after continue to provide warning message.By contrast, if we adopt traditional Principal Component Analysis Model to monitor, Monitoring Performance will reduce greatly, as shown in Figure 4.This has illustrated that also we introduce one-class support vector machines when second layer data modeling be suitable, also is effective.

Table 1: level of factory process monitoring variable list

Above-described embodiment is used for the present invention that explains, rather than limits the invention, and in the protection domain of spirit of the present invention and claim, any modification and change to the present invention makes all fall into protection scope of the present invention.

Claims (6)

1. the level of factory Monitoring of Chemical method that double-deck data model drives is characterized in that, may further comprise the steps:

(1) data of utilizing Distributed Control System (DCS) to collect the level of factory chemical process form the training data sample set that modeling is used: X ∈ R ^{N * m}Wherein, n is the number of sample data collection, and m is the variable number of sample data collection.

(2) with whole level of factory process data collection X ∈ R ^{N * m}Be divided into different submodules

B=1 wherein, 2 ..., B is the sequence number of submodule, B is total submodule number.Deposit these submodule data sets in database backup.

(3) for data set corresponding to each submodule, it is carried out normalized, set up the data statistics monitoring model based on pivot analysis, extract pivot information, set up monitoring and statistics amount and the monitoring limit thereof of submodule.

(4) obtain after the pivot information of each submodule it being re-started arrangement, be combined into new data set and be used for making up second layer data statistic analysis model.

(5) for the new data matrix that mixes pivot information, adopt one-class support vector machines that it is carried out modeling, set up the relational model between each submodule, finish whole modeling and the monitoring of level of factory process.

(6) collect new process data, and it is carried out pre-service and normalization.

(7) new process data is divided into a plurality of submodules, is input in the submodule principal component model, calculate corresponding pivot composition.

(8) reconfigure the pivot information of different submodules, new data matrix is input in the one-class support vector machines model, calculate corresponding monitoring and statistics value, judge the running status of active procedure.

2. the level of factory Monitoring of Chemical method that drives of described double-deck data model according to claim 1 is characterized in that, described step (2) is specially: with whole data set X ∈ R ^{N * m}According to different operating units and functional attributes, be divided into each submodule, that is:

X=[X ₁X ₂…X _b…X _B]，

B=1 wherein, 2 ..., B is the sequence number of submodule, B is total submodule number.

3. the level of factory Monitoring of Chemical method that drives of described double-deck data model according to claim 1, it is characterized in that, described step (3) is specially: for data set corresponding to each submodule, at first it is carried out normalized, namely remove average and the variance of each data variable, then set up Principal Component Analysis Model, specifically can pass through covariance matrix

Carrying out Eigenvalues Decomposition realizes.By pivot analysis, can be divided into principal component space and residual error space, the loading matrix that obtains to former space

And score matrix

As follows:

T _b=X _b·P _b;

Wherein, k _bBe the pivot number of choosing, T _bBe the score matrix of principal component space,

Be the score matrix in residual error space, Decompose the eigenvalue matrix that obtains for pivot, and satisfy following relation Principal Component Analysis Model just is divided into two sub spaces, i.e. principal component space to original process operation space like this

With the residual error space

4. the level of factory Monitoring of Chemical method that drives of described double-deck data model according to claim 1 is characterized in that, described step (4) is specially: the pivot information that obtains each submodule

B=1,2 ..., after the B, it is re-started arrangement, be combined into new data set T _ComBe used for making up second layer data statistic analysis model, that is:

T _com=[T ₁T ₂…T _b…T _B]。

5. the level of factory Monitoring of Chemical method that drives of described double-deck data model according to claim 1 is characterized in that, described step (5) is specially: for the new data matrix T of mixing pivot information _Com, adopt one-class support vector machines that it is carried out modeling, set up the relational model between each submodule, finish whole modeling and the monitoring of level of factory process.At first utilize nonlinear function with T _ComIn each data projection in high-dimensional feature space, namely

Wherein, ti is T _ComIn a data vector, one-class support vector machines is set up model by finding the solution following optimal problem:

Wherein, R and a are respectively radius and the centre ofs sphere of suprasphere in the high-dimensional feature space, Φ (t _i) be the non-linear projection function, C is the one-class support vector machines parameter, and by this parameter, one-class support vector machines can average out between the mistake minute rate of the volume of suprasphere and sample, and ξ is the slack variable of each sample.In the solution procedure of reality, usually adopt following dual propositions to construct one-class support vector machines, namely

Wherein, K (t _i, t _j)=＜Φ (t _i), Φ (t _j) be kernel function, and usually being chosen for the form of gaussian kernel, α is Lagrange multiplier corresponding to each sample.The modeling result of one-class support vector machines is: the α value that most of sample is corresponding is zero, only has α value corresponding to the crucial sample of fraction non-vanishing, and these samples are called as support vector.

6. the level of factory Monitoring of Chemical method that drives of described double-deck data model according to claim 1 is characterized in that, described step (7) and (8) are specially: with new process data x ^NewBe divided into a plurality of submodules, be input in the submodule principal component model, calculate corresponding pivot composition, that is:

On this basis, with the new pivot blended data vector that obtains

Be input in the one-class support vector machines model statistic of using as process monitoring apart from the factor that is defined as follows:

Wherein, D _LimBe the statistics limit of monitoring and statistics amount D, and the radius of suprasphere equates.

Images