CN101382801B - Optimum soft measuring instrument based on EGA-optimized polymerization of propylene production process and method - Google Patents

Abstract

The invention relates to an optimal soft measurement instrument based on EGA optimization of a propylene polymerization production process, which comprises the propylene polymerization production process, an on-site intelligent instrument, a control station, a DCS database for storing data, an optimal soft measurement model based on EGA optimization and a melt index soft measurement value indicator; the on-site intelligent instrument and the control station are connected with the propylene polymerization production process and the DCS database; and the optimal soft measurement model is connected with the DCS database and the soft measurement value indicator. The optimal soft measurement model based on EGA optimization comprises a data preprocessing module, an ICA independent component analyzing module, a BP neural network modeling module and an EGA optimizing module. The invention also provides a soft measurement method realized by the soft measurement instrument. The soft measurement instrument and the soft measurement method realize online measurement, automatic online parameter optimization, high calculation speed, automatic updating of the model, strong anti-interference capability and high precision.

Description

Propylene polymerization production process optimal soft survey instrument and method based on EGA optimization

Technical field

The present invention relates to optimal soft survey instrument and method, specifically is a kind of propylene polymerization production process optimal soft survey instrument and method of optimizing based on EGA.

Background technology

Polypropylene is a kind of thermoplastic resin that is made by propylene polymerization, the most important downstream product of propylene, and 50% of world's propylene, 65% of China's propylene all is to be used for making polypropylene, is one of five big general-purpose plastics, and is closely related with our daily life.Polypropylene is fastest-rising in the world interchangeable heat plastic resin, and total amount only is only second to tygon and Polyvinylchloride.For making China's polypropylene product have the market competitiveness, exploitation rigidity, toughness, crushing-resistant copolymerization product, random copolymerization product, BOPP and CPP film material, fiber, nonwoven cloth that mobile balance is good, and the exploitation polypropylene is in the application of automobile and field of household appliances, all is important research project from now on.

Melting index is that polypropylene product is determined one of important quality index of product grade, it has determined the different purposes of product, measurement to melting index is an important step of production quality control during polypropylene is produced, and to producing and scientific research, important effect and directive significance is arranged all.

Yet; the on-line analysis of melting index is measured and is difficult at present accomplish; being the shortage of online melting index analyser on the one hand, is the conventional online analyser owing to stop up to measure through regular meeting and forbidden even can't normally use difficulty in the use that is caused on the other hand.Therefore, the measurement of MI in the commercial production at present mainly obtains by hand sampling, off-line assay, and can only analyze once in general every 2-4 hour, time lag is big, has brought difficulty to the quality control of propylene polymerization production, becomes to be badly in need of a bottleneck problem solving in the production.The online soft sensor instrument of polypropylene melt index and method research, thus the forward position and the focus of academia and industry member become.

Summary of the invention

Not high for the measuring accuracy that overcomes existing propylene polymerization production process, as to be subject to artificial factor deficiency, propylene polymerization production process melting index optimal soft survey instrument and the method for the invention provides a kind of on-line measurement, on-line parameter Automatic Optimal, computing velocity is fast, model upgrades automatically, antijamming capability is strong, precision is high optimizing based on EGA.

The technical solution adopted for the present invention to solve the technical problems is:

A kind of propylene polymerization production process optimal soft survey instrument of optimizing based on EGA, comprise the field intelligent instrument that is used to measure easy survey variable, the control station that is used for the measuring operation variable, the DCS database of store data and melt index flexible measured value display instrument, described field intelligent instrument, control station is connected with propylene polymerization production process, described field intelligent instrument, control station is connected with the DCS database, described soft measuring instrument also comprises the optimal soft measurement model of optimizing based on EGA, described DCS database is connected with the input end of the described optimal soft measurement model of optimizing based on EGA, the output terminal of the described optimal soft measurement model of optimizing based on EGA is connected with melt index flexible measured value display instrument, and described optimal soft measurement model based on EGA optimization comprises:

Data preprocessing module is used for to the input variable centralization, promptly deducting the mean value of variable with carrying out pre-service from the model input variable of DCS database input; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

ICA independent component analysis module is used for comprising from recovering basic source signal through the pretreated linear hybrid data of data:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ ‖ B ⁺‖;

In the formula, ‖. ‖ represents norm;

(3.4) judge whether convergence, ‖ B ⁺-B ‖＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, establishes output layer k the neuronic actual y of being output as of BP neural network _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

$\mathrm{ne}{t}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function usually, is expressed as:

$f\left(\mathrm{ne}{t}_k\right)=1/(1+e^{-(\mathrm{ne}{t}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if the desired output O of neuron k in the output layer should be arranged mutually _Pk, then the output variance of output layer is expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study; So w _KjCorrection be:

${\mathrm{\Λ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{\∂E_p}{\∂w_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

The EGA optimal module is used to adopt wheel disc to change method as selecting operator, and selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1) (13)

child2＝parent2+α×(parent1-parent2) (14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula.

As preferred a kind of scheme, the described optimal soft measurement model of optimizing based on EGA also comprises: the model modification module, be used for the online updating of model, and regularly the off-line analysis data is input in the training set, upgrade neural network model.

As preferred another scheme: in described data preprocessing module, adopt principal component analytical method to realize that prewhitening handles.

A kind of propylene polymerization production process optimal soft measuring method of optimizing based on EGA, described flexible measurement method mainly may further comprise the steps:

1), to the propylene polymerization production process object, according to industrial analysis and Operations Analyst, selection operation variable and easily survey the input of variable, performance variable and easily survey variable and obtain by the DCS database as model;

2), sample data is carried out pre-service,, promptly deduct the mean value of variable to the input variable centralization; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

3), to carrying out independent component analysis through pretreated data, comprising:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ ‖ B ⁺‖;

In the formula, ‖. ‖ represents norm;

(3.4) judge whether convergence, ‖ B ⁺-B ‖＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, minimizes by error function and finishes a kind of height Nonlinear Mapping that is input to output, keeps topological invariance in the mapping;

4), set up initial neural network model, adopt the BP neural network, establish output layer k the neuronic actual y of being output as of BP neural network based on model input, output data _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function usually, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if the desired output O of neuron k in the output layer should be arranged mutually _Pk, then the output variance of output layer is expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study; So w _KjCorrection be:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

5), adopt the weighting parameter of the initial neural network of EGA optimization method optimization, adopt wheel disc to change method as selecting operator, selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1) (13)

child2＝parent2+α×(parent1-parent2) (14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], and the value of gene σ i will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula.

As preferred a kind of scheme: described flexible measurement method is further comprising the steps of: 6), regularly the off-line analysis data is input in the training set, upgrade neural network model.

Further, in described step 2) in, adopt principal component analytical method to realize the prewhitening processing.

Technical conceive of the present invention is: the important quality index melting index of propylene polymerization production process is carried out the soft measurement of online optimum, overcome the deficiency that existing polypropylene melting index measurement instrument measuring accuracy is not high, be subject to artificial factor, introduce the EGA optimization method neural network parameter and structure are carried out Automatic Optimal, do not need artificial experience or repeatedly test adjust neural network, just can obtain optimum soft measurement result.

Beneficial effect of the present invention mainly shows: 1, on-line measurement; 2, on-line parameter Automatic Optimal; 3, computing velocity is fast; 4, model upgrades automatically; 5, antijamming capability is strong; 6, precision height.

Description of drawings

Fig. 1 is based on the propylene polymerization production process optimal soft survey instrument of EGA optimization and the basic structure synoptic diagram of method;

Fig. 2 is based on the optimal soft measurement model structural representation that EGA optimizes;

Fig. 3 is a propylene polymerization production process Hypol explained hereafter process flow diagram.

Embodiment

Below in conjunction with accompanying drawing the present invention is further described.The embodiment of the invention 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.

Embodiment 1

With reference to Fig. 1, Fig. 2 and Fig. 3, a kind of propylene polymerization production process optimal soft survey instrument of optimizing based on EGA, comprise propylene polymerization production process 1, be used to measure the field intelligent instrument 2 of easy survey variable, the control station 3 that is used for the measuring operation variable, the DCS database 4 of store data and melt index flexible measured value display instrument 6, described field intelligent instrument 2, control station 3 is connected with propylene polymerization production process 1, described field intelligent instrument 2, control station 3 is connected with DCS database 4, described soft measuring instrument also comprises the optimal soft measurement model 5 that EGA optimizes, described DCS database 4 is connected with the input end of the described optimal soft measurement model of optimizing based on EGA 5, the output terminal of the described optimal soft measurement model of optimizing based on EGA 5 is connected with melt index flexible measured value display instrument 6, and described optimal soft measurement model based on EGA optimization comprises:

Data preprocessing module is used for to the input variable centralization, promptly deducting the mean value of variable with carrying out pre-service from the model input variable of DCS database input; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

ICA independent component analysis module is used for comprising from recovering basic source signal through the pretreated linear hybrid data of data:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ ‖ B ⁺‖;

In the formula, ‖. ‖ represents norm;

(3.4) judge whether convergence, ‖ B ⁺-B ‖ ‖＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, establishes output layer k the neuronic actual y of being output as of BP neural network _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function usually, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if the desired output O of neuron k in the output layer should be arranged mutually _Pk, then the output variance of output layer is expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study; So w _KjCorrection be:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

The EGA optimal module is used to adopt wheel disc to change method as selecting operator, and selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1) (13)

child2＝parent2+α×(parent1-parent2) (14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula.

The described optimal soft measurement model of optimizing based on EGA also comprises: the model modification module, be used for the online updating of model, and regularly the off-line analysis data is input in the training set, upgrade neural network model.

In described data preprocessing module, adopt principal component analytical method to realize the prewhitening processing.

The propylene polymerization production process process flow diagram as shown in Figure 3, according to reaction mechanism and flow process analysis, consider the various factors that in the polypropylene production process melting index is exerted an influence, get nine performance variables commonly used in the actual production process and easily survey variable as the modeling variable, have: three strand of third rare feed flow rates, major catalyst flow rate, cocatalyst flow rate, hydrogen volume concentration in temperature in the kettle, pressure, the liquid level, still.

The required modeling variable of optimal soft measurement model that table 1 is optimized based on EGA

Table 1 has been listed 9 modeling variablees as optimal soft measurement model 5 inputs of optimizing based on EGA, is respectively liquid level (L) in temperature in the kettle (T), still internal pressure (p), the still, the interior hydrogen volume concentration (X of still _v), 3 bursts of propylene feed flow rates (first strand of third rare feed flow rates f1, second strand of third rare feed flow rates f2, the 3rd strand of third rare feed flow rates f3), 2 bursts of catalyst charge flow rates (major catalyst flow rate f4, cocatalyst flow rate f5).Polyreaction in the reactor is that reaction mass mixes back participation reaction repeatedly, so the model input variable relates to the mean value in preceding some moment of process variable employing of material.Last hour mean value of The data in this example.Melting index off-line laboratory values is as the output variable of the optimal soft measurement model of optimizing based on EGA 5.Obtain by hand sampling, off-line assay, analyzed collection once in per 4 hours.

Field intelligent instrument 2 and control station 3 link to each other with propylene polymerization production process 1, link to each other with DCS database 4; Optimal soft measurement model 5 links to each other with DCS database and soft measured value display instrument 6.Field intelligent instrument 2 is measured the easy survey variable that propylene polymerization is produced object, will easily survey variable and be transferred to DCS database 4; Control station 3 control propylene polymerizations are produced the performance variable of object, and performance variable is transferred to DCS database 4.The variable data of record is as the input of the optimal soft measurement model of optimizing based on EGA 5 in the DCS database 4, and soft measured value display instrument 6 is used to show the output of the optimal soft measurement model of optimizing based on EGA 5, promptly soft measured value.

Optimal soft measurement model 5 based on EGA optimizes comprises:

Data preprocessing module 7 is used for model input carrying out pre-service, i.e. centralization and prewhitening.To the input variable centralization, deduct the mean value of variable exactly, making variable is the variable of zero-mean, thus shortcut calculation.It is the variable decorrelation that the input variable prewhitening is handled, and input variable is applied a linear transformation, makes between each component of variable after the conversion uncorrelatedly mutually, and its covariance matrix is a unit matrix simultaneously.Generally realize by principal component analytical method.

ICA independent component analysis module 8 is from through recovering the method for basic source signal the data pretreated linear hybrid data.Being described below of ICA problem:

Suppose to have n observational variable x1, x2 ..., xn, they are independent component variable s1 of m non-Gaussian distribution, s2 ..., the linear combination of sm.Contextual definition between the two is:

X＝AS+N (1)

In the formula, A is unknown hybrid matrix, and N is the observation noise vector.X＝[x1，x2，...，xn]，S＝[s1，s2，...，sm]。

Following formula is the ICA basic model, and the expression observation data is how to mix generation by the independent component component.Independent component is implicit variable, mean and can not directly observe, and mixing coefficient matrix A also is unknown that known only is observational variable X, how to utilize observational variable X to estimate A and S, just the ICA problem that will solve.The purpose of ICA will be sought exactly and separate mixed matrix B, can obtain separate source variable by observational variable X by it:

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

Concrete steps are as follows:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of matrix X;

(3.3) allow B=B ⁺/ ‖ B ⁺‖;

In the formula, ‖. ‖ represents norm;

(3.4) judge whether convergence, ‖ B ⁺-B ‖＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B.

Then, calculate each separation component by formula (2).

Neural net model establishing module 9 is implemented as follows:

Adopt the BP neural network, multilayer feedforward neural network is made up of input layer, hidden layer and output layer on network structure usually.On network characterization, mainly show as both do not had the layer in neuronic interconnected, do not have the anti-contact of interlayer yet.This network comes down to a kind of static network, and its output is the function of existing input, and irrelevant with inputing or outputing of past and future.By the BP neural network structure as can be known, arbitrary neuronic weighted sum that is output as the input pattern component in the input layer, the notion of this weighted sum are fit to all the other each layers equally.If the output layer k of BP neural network neuronic actual y that is output as _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

yk＝f(netk) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function usually, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function.Make that training sample is k.For any input pattern X _p, if k neuronic desired output O in the output layer should be arranged mutually _Pk, then the output variance of output layer can be expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output.Oppositely error propagation the destination of study is to revise connection weight w value, makes E reach minimum value.This just requires connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study.So w _KjCorrection be:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor.

4) the EGA optimal module 10, are implemented as follows:

Select operator to adopt wheel disc to change method, selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1) (13)

child2＝parent2+α×(parent1-parent2)?(14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces.Than traditional arithmetic crossover operator, offspring individual no longer is confined to overcome the search volume and constantly shunk on the line of parent individuality, precocious easily shortcoming.

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula.Model modification module 11 is used for the online updating of model, regularly the off-line analysis data is input in the training set, upgrades neural network model.

Embodiment 2

With reference to Fig. 1, Fig. 2 and Fig. 3, a kind of propylene polymerization production process optimal soft measuring method of optimizing based on EGA, described flexible measurement method mainly may further comprise the steps:

1), to the propylene polymerization production process object, according to industrial analysis and Operations Analyst, selection operation variable and easily survey the input of variable, performance variable and easily survey variable and obtain by the DCS database as model;

2), sample data is carried out pre-service,, promptly deduct the mean value of variable to the input variable centralization; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

3), to carrying out independent component analysis through pretreated data, comprising:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ ‖ B ⁺‖;

In the formula, ‖. ‖ represents norm;

(3.4) judge whether convergence, ‖ B ⁺One B ‖＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, minimizes by error function and finishes a kind of height Nonlinear Mapping that is input to output, keeps topological invariance in the mapping;

4), set up initial neural network model, adopt the BP neural network, establish output layer k the neuronic actual y of being output as of BP neural network based on model input, output data _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function usually, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if k neuronic desired output O in the output layer should be arranged mutually _Pk, then the output variance of output layer is expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study; So w _KjCorrection be:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

5), adopt the weighting parameter of the initial neural network of EGA optimization method optimization, adopt wheel disc to change method as selecting operator, selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1) (13)

child2＝parent2+α×(parent1-parent2) (14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula.

Further, in described step 2) in, adopt principal component analytical method to realize the prewhitening processing.

The concrete implementation step of the method for present embodiment is as follows:

Step 1: to propylene polymerization production process object 1, according to industrial analysis and Operations Analyst, selection operation variable and easy input of surveying variable as model.Performance variable and the easy variable of surveying are obtained by DCS database 4.

Step 2: sample data is carried out pre-service, finish by data preprocessing module 7.

Step 3:, finish by ICA independent component analysis module 8 to carrying out independent component analysis through pretreated data.

Step 4: set up initial neural network model 9 based on model input, output data.The input data are as acquisition as described in the step 1, and output data is obtained by the off-line chemical examination.

Step 5: the weighting parameter of optimizing initial neural network 8 by EGA optimization method 10.

Step 6: model modification module 11 regularly is input to the off-line analysis data in the training set, upgrades neural network model, optimizes optimal soft measurement model 5 foundation of BP neural network based on EGA and finishes.

Step 7: melt index flexible measured value display instrument 6 shows the output of optimizing the optimal soft measurement model 5 of BP neural network based on EGA, finishes the demonstration to the optimum soft measurement of propylene polymerization production process melting index.

Claims (6)

1. propylene polymerization production process optimal soft survey instrument of optimizing based on EGA, comprise the field intelligent instrument that is used to measure easy survey variable, the control station that is used for the measuring operation variable, the DCS database of store data and melt index flexible measured value display instrument, described field intelligent instrument, control station is connected with propylene polymerization production process, described field intelligent instrument, control station is connected with the DCS database, it is characterized in that: described soft measuring instrument also comprises the optimal soft measurement model of optimizing based on EGA, described DCS database is connected with the input end of the described optimal soft measurement model of optimizing based on EGA, the output terminal of the described optimal soft measurement model of optimizing based on EGA is connected with melt index flexible measured value display instrument, the described optimal soft measurement model of optimizing based on EGA comprises: data preprocessing module, be used for and carry out pre-service from the model input variable of DCS database input, to the input variable centralization, promptly deduct the mean value of variable; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

ICA independent component analysis module is used for comprising from recovering basic source signal through the pretreated linear hybrid data of data:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ || B ⁺||;

In the formula, || .|| represents norm;

(3.4) judge whether convergence, || B ⁺-B||＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, establishes the actual y of being output as of the output layer neuron k of BP neural network _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if k neuronic desired output O in the output layer should be arranged mutually _Pk, the output variance E of output layer then _pBe expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study; So w _KjCorrection amount _pw _KjFor:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

The EGA optimal module is used to adopt wheel disc to change method as selecting operator, and selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1)(13)

child2＝parent2+α×(parent1-parent2)(14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula, σ _i' be by the gene after the variation.

2. the propylene polymerization production process optimal soft survey instrument of optimizing based on EGA as claimed in claim 1 is characterized in that: the described optimal soft measurement model of optimizing based on EGA also comprises:

The model modification module is used for the online updating of model, regularly the off-line analysis data is input in the training set, upgrades neural network model.

3. the propylene polymerization production process optimal soft survey instrument of optimizing based on EGA as claimed in claim 1 or 2 is characterized in that: in described data preprocessing module, adopt principal component analytical method to realize the prewhitening processing.

4. flexible measurement method of realizing with the propylene polymerization production process optimal soft survey instrument of optimizing based on EGA as claimed in claim 1, it is characterized in that: described flexible measurement method mainly may further comprise the steps:

1), to the propylene polymerization production process object, according to industrial analysis and Operations Analyst, selection operation variable and easily survey the input of variable, performance variable and easily survey variable and obtain by the DCS database as model;

2), the model input variable of DCS database input is carried out pre-service,, promptly deduct the mean value of variable to the input variable centralization; The input variable prewhitening being handled is the variable decorrelation again, and input variable is applied a linear transformation;

3), to carrying out independent component analysis through pretreated data, comprising:

(3.1) select initial weight B at random;

(3.2) B is carried out iteration and upgrade B ⁺=E{xg (B ^TX) }-E{g ' (B ^TX) } B;

In the formula, B ⁺Weights after expression is upgraded, B ^TBe the transposition of B, g ' is the derivative of g, and E is a mathematical expectation, and x is the vector of observational variable matrix X;

(3.3) allow B=B ⁺/ || B ⁺||;

In the formula, || .|| represents norm;

(3.4) judge whether convergence, || B ⁺-B||＜epsilon (6)

Epsilon represents to restrain index in the formula, does not restrain and returns (3.2), otherwise continue;

(3.5) storage B;

Then, calculate each separation component by formula (2):

Y＝BX (2)

Y is the estimated vector of S in the formula, and S is the independent component matrix of variables, and X is the observational variable matrix;

Estimating of separating resulting independence adopts the independence criterion based on negentropy

J(y)∝[E{G(y)}-E{G(y)} ²] (3)

G () is non-quadratic function in the formula, and y is the vector of matrix Y; Select:

G(y)＝-exp(-y ²/2) (4)

Formula (3) estimates based on the negentropy of entropy principle that promptly when negentropy J (y) was maximum, variable was independent;

Estimate separation matrix B, iterate based on point of fixity and seek BX,, get to be non-Gauss's maximization of criterion based on negentropy formula (3):

g(x)＝xexp(-x ²/2) (5)；

Function g () is the derivative of function G () in the formula;

The neural net model establishing module is used to adopt the BP neural network, minimizes by error function and finishes a kind of Nonlinear Mapping that is input to output, keeps topological invariance in the mapping;

4), set up initial neural network model, adopt the BP neural network, establish the actual y of being output as of the output layer neuron k of BP neural network based on model input, output data _k, be input as net _k, arbitrary neuron j is output as y in the hidden layer that layer is adjacent therewith _j, then have:

${\mathrm{net}}_k=\underset{i}{\mathrm{\Σ}}{w}_{\mathrm{kj}}\·y_j---\left(7\right)$

y _k＝f(net _k) (8)

In the formula, w _KjBe the connection weight between neuron k and the neuron j, f () is neuronic output function, is taken as the Sigmoid function, is expressed as:

$f\left({\mathrm{net}}_k\right)=1/(1+e^{-({\mathrm{net}}_k+h_k)/{\mathrm{\θ}}_0})---\left(9\right)$

In the formula, h _kBe the threshold value of neuron k, θ ₀Be the steepness parameter, in order to regulate the steepness of Sigmoid function;

Make that training sample is K, for any input pattern X _p, if the desired output O of neuron k in the output layer should be arranged mutually _Pk, the output variance E of output layer then _pBe expressed as:

$E_p=\frac{1}{2}\underset{p}{\mathrm{\Σ}}{(O_{\mathrm{pk}}-y_{\mathrm{pk}})}^2---\left(10\right)$

In the formula, O _PkRepresent desired output, y _PkRepresent actual output; Oppositely error propagation the destination of study is to revise connection weight w value, makes E _pReach minimum value; Require connection weight w _Kj, w _JiShould be along E _pNegative gradient direction study;

So w _KjCorrection amount _pw _KjFor:

${\mathrm{\Δ}}_p{w}_{\mathrm{kj}}=-\mathrm{\β}\frac{{\∂E}_p}{{\∂w}_{\mathrm{kj}}}---\left(11\right)$

In the formula, β is that learning rate is adjusted the factor;

5), adopt the weighting parameter of the initial neural network of EGA optimization method optimization, adopt wheel disc to change method as selecting operator, selective rule is as follows: establishing group size is n, and the fitness of individual i is f _i, the selecteed probability P of then individual i is expressed as:

$P=f_i/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{f}_j---\left(12\right)$

The Crossover Strategy that crossover operator adopts arithmetic to intersect, and on the basis of conventional cross operator, done following change:

child1＝parent1+α×(parent2-parent1)(13)

child2＝parent2+α×(parent1-parent2)(14)

α is the random number on [0.25,1.25] interval in the formula, and parent1, parent2 are a parent voxel vector, and child1, child2 are for intersecting the offspring individual vector that produces;

Mutation operator adopts evenly variation, establishes individual span and is [a, b], gene σ _iValue will be by variation:

σ _i′＝a+γ×(b-a) (15)

γ is the random number on [0,1] in the formula, σ _i' be by the gene after the variation.

5. flexible measurement method as claimed in claim 4 is characterized in that: described flexible measurement method is further comprising the steps of: 6), regularly the off-line analysis data is input in the training set, upgrade neural network model.

6. as claim 4 or 5 described flexible measurement methods, it is characterized in that: in described step 2) in, adopt principal component analytical method to realize the prewhitening processing.

Images