CN103823430A - Intelligent weighing propylene polymerization production process optimal soft measurement system and method - Google Patents

Abstract

The invention discloses an intelligent weighing propylene polymerization production process optimal soft measurement system which comprises a propylene polymerization production process, a field intelligent instrument, a control station, a DCS database which stores data, an intelligent weighing optimal soft measurement system and a melt index soft measurement value display instrument. The field intelligent instrument and the control station are connected with the propylene polymerization production process, and are connected with the DCS database. The optimal soft measurement system is connected with the DCS database and the soft measurement value display instrument. The intelligent weighing optimal soft measurement system comprises a model updating module, a data preprocessing module, a PCA principal component analysis module, a neural network model module, an intelligent weighing optimization module and a model updating module. The invention further provides a soft measurement method which is realized by using the soft measurement system. According to the invention, online measurement and online parameter optimization are realized, and the method and the system have the advantages of quick soft measurement, automatic model updating, strong anti-interference ability and high precision.

Description

Intelligence weighting propylene polymerization production process optimal soft measuring system and method

Technical field

The present invention relates to a kind of optimum soft measuring system and method, specifically a kind of intelligent weighting propylene polymerization production process optimal soft measuring system and method.

Background technology

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

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

But; 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 that existing in-line analyzer is measured the inaccurate difficulty in caused use that even cannot normally use owing to often can stopping up on the other hand.Therefore, the measurement of MI in commercial production at present, is mainly to obtain by hand sampling, off-line assay, and can only analyze once for general every 2-4 hour, time lag is large, and the quality control of producing to propylene polymerization has brought difficulty, becomes a bottleneck problem being badly in need of solution in production.The online soft sensor system and method research of polypropylene melt index, thus forward position and the focus of academia and industry member become.

Summary of the invention

In order to overcome, the measuring accuracy of current existing propylene polymerization production process is not high, the deficiency of the impact that is subject to human factor, the object of the present invention is to provide a kind of on-line measurement, on-line parameter optimization, soft measuring speed is fast, model upgrades automatically, antijamming capability is strong, precision the is high optimum soft measuring system of intelligent weighting propylene polymerization production process melting index and method.

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

A kind of intelligent weighting propylene polymerization production process optimal soft measuring system, comprise propylene polymerization production process, for measuring the field intelligent instrument of easy survey variable, for measuring the control station of performance variable, the DCS database of store data, the intelligence soft measuring system of Weighted optimal and melt index flexible are measured display instrument, described field intelligent instrument, control station is connected with propylene polymerization production process, described field intelligent instrument, control station is connected with DCS database, described DCS database is connected with the input end of the soft measuring system of intelligent Weighted optimal, the output terminal of the soft measuring system of described intelligent Weighted optimal is measured display instrument with melt index flexible and is connected, it is characterized in that: the soft measuring system of described intelligent Weighted optimal comprises:

(1), data preprocessing module, for the mode input variable from DCS database input is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(2), PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realize by input variable being applied to linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(3), neural network model module, for adopting RBF neural network, minimized to be input to a kind of nonlinear of output by error function, in mapping, keep topological invariance; Need to set up some sub neural networks, the training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">J</figure-callout>}}_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(4), intelligent weighted optimization module, for each sub-network of step (3) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{\mathrm{<figure-callout\; id="1"\; label="I"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">I</figure-callout>}}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

As preferred a kind of scheme, described intelligent Weighted optimal soft-sensing model also comprises: model modification module, for the online updating of model, will regularly off-line analysis data be input in training set, and upgrade neural network model.

As preferred another scheme: in described intelligent continuous space ant group algorithm training weighting multimode RBF neural network model, train sub-RBF neural network, then its weighted array is got up to form neural network group; Because the selection standard of sub-network is that prediction error is little, large with other sub-network difference, so these values of forecasting are good, the comprehensive forecasting effect of different sub neural network can have better forecast precision and stability again.

As preferred another scheme: in PCA principal component analysis (PCA) module, PCA method realizes the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improves the performance of model.

The flexible measurement method that the optimum soft measuring system of intelligent weighting polypropylene production process realizes, described flexible measurement method specific implementation step is as follows:

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

(2) sample data is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(3) PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realizes by input variable being applied to a linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(4) several the initial sub neural network models of setting up based on mode input, output data intelligence continuous space ant group algorithm training smart, adopt RBF neural network, complete a kind of nonlinear that is input to output by error minimize, in mapping, keep topological invariance; The training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">J</figure-callout>}}_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(5), all sub neural networks of weighted array, for each sub-network of step (5.4) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

Technical conceive of the present invention is: the important quality index melting index of propylene polymerization production process is carried out to online optimum soft measurement, overcome that existing polypropylene melt index measurement instrument measuring accuracy is not high, the deficiency of the impact that is subject to human factor, set up by the method for intelligent continuous space ant group algorithm training weighting multimode RBF neural network that forecast precision is high, the forecasting model of good stability obtains optimum soft measurement result.

Beneficial effect of the present invention is mainly manifested in: 1, on-line measurement; 2, on-line parameter Automatic Optimal; 3, soft measuring speed is fast; 4, model upgrades automatically; 5, antijamming capability is strong; 6, precision is high.

Accompanying drawing explanation

Fig. 1 is the basic structure schematic diagram of intelligent weighting propylene polymerization production process optimal soft measuring system and method;

Fig. 2 is the soft measuring system structural representation of intelligent Weighted optimal;

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

Embodiment

Below in conjunction with accompanying drawing, the present invention is described further.The embodiment of the present 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 that the present invention is made, all fall into protection scope of the present invention.

Embodiment 1

1. with reference to Fig. 1, Fig. 2 and Fig. 3, a kind of intelligent weighting propylene polymerization production process optimal soft measuring system, comprise propylene polymerization production process 1, for measuring the field intelligent instrument 2 of easy survey variable, for measuring the control station 3 of performance variable, the DCS database 4 of store data, the intelligence soft measuring system 5 of Weighted optimal 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 DCS database 4 is connected with the input end of the soft measuring system 5 of intelligent Weighted optimal, the output terminal of the soft measuring system 5 of described intelligent Weighted optimal is connected with melt index flexible measured value display instrument 6, the soft measuring system of described intelligent Weighted optimal comprises:

(1), data preprocessing module, for the mode input variable from DCS database input is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(2), PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realize by input variable being applied to linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(3), neural network model module, for adopting RBF neural network, minimized to be input to a kind of nonlinear of output by error function, in mapping, keep topological invariance; Need to set up some sub neural networks, the training objective of first sub-RBF network is forecast result and actual result gap J1 minimum;

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">J</figure-callout>}}_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(4), intelligent weighted optimization module, for each sub-network of step (3) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

In PCA principal component analysis (PCA) module, PCA method realizes the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improves the performance of model.

2. propylene polymerization production process process flow diagram as shown in Figure 3, according to reaction mechanism and flow process analysis, consider each factor in polypropylene production process, melting index being exerted an influence, get nine performance variables conventional in actual production process and easily survey variable as mode input variable, have: three bursts of propylene feed flow rates, major catalyst flow rate, cocatalyst flow rate, temperature in the kettle, pressure, liquid level, hydrogen volume concentration in still.

Table 1 has been listed 9 mode input variablees inputting as the soft measuring system 5 of intelligent Weighted optimal, is respectively hydrogen volume concentration (X in pressure (p) in temperature in the kettle (T), still, the interior liquid level (L) of still, still _v), 3 bursts of propylene feed flow rates (first gang of propylene feed flow rate f1, second gang of propylene feed flow rate f2, the 3rd gang of propylene feed flow rate f3), 2 bursts of catalyst charge flow rates (major catalyst flow rate f4, cocatalyst flow rate f5).Polyreaction in reactor is that reaction mass mixes rear participation reaction repeatedly, and therefore mode input variable relates to the mean value in front some moment of process variable employing of material.The mean value of last hour for data acquisition in this example.Melting index off-line laboratory values is as the output variable of the soft measuring system 5 of intelligent Weighted optimal.Obtain by hand sampling, off-line assay, within every 4 hours, analyze and gather once.

The required mode input variable of the soft measuring system of the intelligent Weighted optimal of table 1

Field intelligent instrument 2 and control station 3 are connected with propylene polymerization production process 1, are connected with DCS database 4; Optimum soft measuring system 5 is connected with DCS database 4 and soft measured value display instrument 6.Field intelligent instrument 2 is measured the easy survey variable of propylene polymerization production object, will easily survey variable and be transferred to DCS database 4; Control station 3 is controlled the performance variable of propylene polymerization production object, and performance variable is transferred to DCS database 4.In DCS database 4, the variable data of record is as the input of the soft measuring system 5 of intelligent Weighted optimal, and soft measured value display instrument 6 is for showing the output of the soft measuring system 5 of intelligent Weighted optimal, i.e. soft measured value.

The soft measuring system 5 of intelligence Weighted optimal, comprising:

(1) data preprocessing module 7, for mode input is carried out to pre-service, i.e. centralization and normalization.To input variable centralization, deduct exactly the mean value of variable, making variable is the variable of zero-mean, thus shortcut calculation; To input variable normalization, be exactly the constant interval divided by input variable value, be that the value of variable is fallen within-0.5～0.5, further simplify.

(2) PCA principal component analysis (PCA) module 8, for to input variable prewhitening, processing is variable decorrelation, input variable is applied to a linear transformation, make between each component of variable after conversion uncorrelated mutually, its covariance matrix is unit matrix simultaneously, and major component is obtained by C=xU, and wherein x is input variable, C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix.

(3) neural network model module 9, adopts RBF neural network, and multilayer feedforward neural network is conventionally made up of input layer, hidden layer and output layer in network structure.On network characterization main manifestations for both without neuronic interconnected in layer, also without the anti-contact of interlayer.This network is in fact a kind of static network, and its output is the function of existing input, and irrelevant in inputing or outputing of past and future.RBF neural network model has an input layer, an output layer and a hidden layer.Can prove in theory, RBF neural network can be approached arbitrarily nonlinear system.RBF neural network BP training algorithm has minimized to be input to a kind of nonlinear of output by error function, keep topological invariance in mapping; Need to set up some sub neural networks, the training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">J</figure-callout>}}_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(4), intelligent weighted optimization module 10, for each sub-network of step (3) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

In PCA principal component analysis (PCA) module, PCA method realizes the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improves the performance of model.

(5) model modification module 11, for the online updating of model, is regularly input to off-line analysis data in training set, upgrades neural network model.

Embodiment 2

1. with reference to Fig. 1, Fig. 2 and Fig. 3, a kind of intelligent weighting propylene polymerization production process optimal soft measuring method comprises the following steps:

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

(2) sample data is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(3) PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realizes by input variable being applied to a linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(4) set up initial neural network model based on mode input, output data, adopt RBF neural network, complete a kind of nonlinear that is input to output by error minimize, in mapping, keep topological invariance; The training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

${\mathrm{<figure-callout\; id="1"\; label="J"\; filenames="HDA0000432564380000011.png"\; state="\{\{state\}\}">J</figure-callout>}}_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(5), all sub neural networks of weighted array, for each sub-network of step (5.4) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

Further, in described step (3), adopt PCA principal component analytical method to realize the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improve the performance of model.

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

Step 1: to propylene polymerization production process object 1, according to industrial analysis and Operations Analyst, select performance variable and easily survey the input of variable as model.

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

Step 3: to carrying out principal component analysis (PCA) through pretreated data, completed by PCA principal component analysis (PCA) module 8.

Step 4: module 9 is set up some initial neural network models based on mode input, output integrating step (4).Input data obtain as described in step 1, and output data are obtained by off-line chemical examination.

Step 5: module 10 integrating steps (5) are combining all sub neural network intelligence according to sub-network prediction error;

Step 6: model modification module 11 is regularly input to off-line analysis data in training set, upgrades neural network model, and the soft measuring system 5 of intelligent Weighted optimal has been set up.

Step 7: the real-time model input variable data that the soft measuring system 5 of intelligent Weighted optimal establishing transmits based on DCS database 4 are carried out the soft measurement of intelligent Weighted optimal to the melting index of propylene polymerization production process 1.

Step 8: melt index flexible is measured the output that display instrument 6 shows the soft measuring system 5 of intelligent Weighted optimal, completes the demonstration of the soft measurement of optimum to propylene polymerization production process melting index.

Claims (2)

1. an intelligent weighting propylene polymerization production process optimal soft measuring system, comprise propylene polymerization production process, for measuring the field intelligent instrument of easy survey variable, for measuring the control station of performance variable, the DCS database of store data, the intelligence soft measuring system of Weighted optimal and melt index flexible are measured display instrument, described field intelligent instrument, control station is connected with propylene polymerization production process, described field intelligent instrument, control station is connected with DCS database, described DCS database is connected with the input end of the soft measuring system of intelligent Weighted optimal, the output terminal of the soft measuring system of described intelligent Weighted optimal is measured display instrument with melt index flexible and is connected, it is characterized in that: the soft measuring system of described intelligent Weighted optimal comprises:

(1), data preprocessing module, for the mode input variable from DCS database input is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(2), PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realize by input variable being applied to linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(3), neural network model module, for adopting RBF neural network, minimized to be input to a kind of nonlinear of output by error function, in mapping, keep topological invariance; Need to set up some sub neural networks, the training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

$J_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$ N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(4), intelligent weighted optimization module, for each sub-network of step (3) is composed to weights; According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

The soft measuring system of described intelligent Weighted optimal also comprises: model modification module, for the online updating of model, regularly off-line analysis data is input in training set, and upgrade neural network model.

The described soft measuring system of intelligent Weighted optimal, is characterized in that: train sub-RBF neural network, then its weighted array is got up to form neural network group; Because the selection standard of sub-network is that prediction error is little, large with other sub-network difference, so these values of forecasting are good, the comprehensive forecasting effect of different sub neural network can have better forecast precision and stability again.In PCA principal component analysis (PCA) module, PCA method realizes the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improves the performance of model.

2. a flexible measurement method of realizing by the optimum soft measuring system of intelligent weighting polypropylene production process as claimed in claim 1, is characterized in that: described flexible measurement method specific implementation step is as follows:

(5.1) to propylene polymerization production process object, according to industrial analysis and Operations Analyst, select performance variable and easily survey the input of variable as model, general operation variable and easily survey variable are got temperature, pressure, liquid level, hydrogen gas phase percentage, 3 strands of propylene feed flow velocitys and 2 strands of these variablees of catalyst charge flow velocity, are obtained by DCS database;

(5.2) sample data is carried out to pre-service, to input variable centralization, deduct the mean value of variable; Be normalized again, divided by the constant interval of variate-value;

(5.3) PCA principal component analysis (PCA) module, for input variable prewhitening is processed and variable decorrelation, realizes by input variable being applied to a linear transformation, be that major component is obtained by C=xU, wherein x is input variable, and C is principal component scores matrix, and U is loading matrix.If raw data is reconstructed, can be by x=CU ^tcalculate the wherein transposition of subscript T representing matrix.In the time that the major component number of choosing is less than the variable number of input variable, x=CU ^t+ E, wherein E is residual matrix;

(5.4) set up several initial sub neural network models based on mode input, output data, adopt RBF neural network, complete a kind of nonlinear that is input to output by error minimize, in mapping, keep topological invariance; The training objective of first sub-RBF network is forecast result and actual result gap J ₁minimum;

$J_1=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_1\left(x_l\right)-d\left(x_l\right))}^2---\left(1\right)$

N is number of samples, and x is input variable, and l is sample point sequence number, F ₁() is sub-network forecast result, and d () is actual result.

Since second sub-network, training objective becomes and makes the prediction error of network as far as possible little, the forecast result of network and network forecast result before large as far as possible difference again simultaneously, and objective function is as follows:

$J_i=\frac{1}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-d\left(x_l\right))}^2-\frac{\mathrm{\&lambda;}}{N}\underset{l=1}{\overset{N}{\mathrm{\&Sigma;}}}{(F_i\left(x_l\right)-F\left(x_l\right))}^2---\left(2\right)$

Ji is the training objective of a front i sub-network, F _i() is the forecast result of i network; D () is actual result; F () is the synthesis result of a front i-1 sub-network; λ is for regulating parameter, and N is number of samples.

The end condition of training is that the new sub-network obtaining is added after multimode neural network, and the prediction error of network group no longer reduces.

Adopt a kind of continuous space ant group algorithm to train and optimization each RBF network, concrete steps are:

(a) algorithm initialization, constructs initial disaggregation S=(s according to RBF neural network structure to be optimized ₁, s ₂..., s _n), the number that n is initial solution, sn is n initial solution, determines ant group's big or small M, and the threshold value MaxGen of ant optimization algorithm iteration number of times the iterations sequence number gen=0 of initialization ant optimization are set;

(b) calculate the fitness value G that disaggregation S is corresponding _i(i=1,2 ..., n), the larger Xie Yuehao that represents of fitness value; Determining to separate according to following formula again concentrates each solution to be got the probability P as the initial solution of ant optimizing _i(i=1,2 ..., n)

$P_a\left(k\right)=\frac{G_a}{{\mathrm{\&Sigma;}}_{a=1}^n{G}_i}(a=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,n)---\left(3\right)$

N is the number of initial solution, and sn is n initial solution, and k is iterations.The ant numbering a=0 of optimizing algorithm is carried out in initialization;

(c) ant a chooses a solution in the S initial solution as optimizing, and selection rule is to do wheel disc choosing according to P;

(d) ant a carries out optimizing on the basis of the initial solution of choosing, and finds better solution s _a';

(e) if a<M, a=a+1, returns to step c; Otherwise continue execution step f downwards;

(f) if gen<MaxGen, gen=gen+1, uses the better solution that all ants obtain in steps d to replace the homographic solution in S, returns to step b; Otherwise perform step g downwards;

(g) calculate the fitness value G that disaggregation S is corresponding _a(a=1,2 ..., n), choose the solution of fitness value maximum as the optimum solution of algorithm, finish algorithm and return.

Each ant fixing number of times that can circulate when optimizing on the basis of its selected initial solution if this circulation has obtained better solution, can and keep the direction of search constant based on this solution in circulation next time; Otherwise the solution based on original but can adjust the direction of search still in next time circulation;

Along with the increase of whole ant optimization algebraically, the step-length of Ant Search can reduce intelligently simultaneously, to be applicable to the convergence of whole ant optimization:

del _k=Random·k ^r (4)

In formula, del _kfor ant k is for the initial step length of iteration, k is iteration algebraically, and Random is random vector, and r is negative normal real number.

For medium-term and long-term those solutions that are not elected to be optimizing initial solution by ant of disaggregation S, can adopt variation and Crossover Strategy in genetic algorithm to process, thereby improve the global optimizing performance of algorithm.

(5.5) all sub neural networks of weighted array, compose weights for the each sub-network to step (5.4); According to being the prediction error of each sub-network, error is less, and weights are larger;

$w_q=(1/e_q)/(\underset{j=1}{\overset{I}{\mathrm{\&Sigma;}}}1/e_j),q=\mathrm{1,2},\&CenterDot;\&CenterDot;\&CenterDot;,I.---\left(5\right)$

$e_j=\frac{1}{N}\underset{m=1}{\overset{N}{\mathrm{\&Sigma;}}}|F_j\left(x_m\right)-d\left(x_m\right)|---\left(6\right)$

Wq is the weights of q sub-network; Eq is the prediction error of q sub-network; I is total sub-network number; J is sub-network sequence number; N is number of samples, and x is input variable, and m is sample point sequence number, F _j() is j sub-network forecast result, and d () is actual result.

The forecast result of final multimode neural network is the weighted sum of each sub-network forecast result.

$O\left(x\right)=\underset{k=1}{\overset{I}{\mathrm{\&Sigma;}}}(w_k\&CenterDot;F_k\left(x\right))---\left(7\right)$

In formula, x is input variable, and O () is model output, F _k() is k sub-network output, w _kbe the weight of k sub-network, I is sub-network sum.

Described flexible measurement method, also comprises: regularly off-line analysis data is input in training set, upgrades neural network model.

Described flexible measurement method, is characterized in that: train sub-RBF neural network, then its weighted array is got up to form multimode neural network; Because the selection standard of sub-network is that prediction error is little, large with other sub-network difference, so these values of forecasting are good, the comprehensive forecasting effect of different sub neural network can have better forecast precision and stability again.In described step (5.3), adopt PCA principal component analytical method to realize the prewhitening processing of input variable, can simplify the input variable of neural network model, and then improve the performance of model.

Images