CN104634478A - Soft measurement method for burning zone temperature of rotary kiln - Google Patents

Abstract

The invention discloses a soft measurement method for the burning zone temperature of a rotary kiln. A large quantity of parameter data, including coal gas flow, rod pushing time, kiln head temperature, kiln tail temperature and the like, can be collected on a working site of the lime rotary kiln. However, the data have different influences on the burning zone temperature. According to the soft measurement method, kernel principal component analysis (KPCA) is introduced to process the data, key parameters influencing the burning zone temperature are searched out for modeling, sample dimension is reduced, and modeling difficulty is lowered; a least square support vector machine (LSSVM) is simple in structure, has a complete theoretical basis and only needs few samples, thereby being quite suitable for soft measurement of the burning zone temperature of the rotary kiln; meanwhile, modeling parameters of the least square support vector machine are optimized by a particle swarm optimization (PSO) method, and modeling and predicating accuracy is improved; the soft measurement method is also applicable to other types of rotary kilns.

Description

A kind of flexible measurement method of calcined by rotary kiln band temperature

Technical field

The present invention relates to a kind of temperature checking method, be specifically related to the hard measurement detection method of lime rotary kiln calcining belt temperature.

Background technology

Lime rotary kiln is the key equipment in metallurgy industry, and its running status directly affects product quality, energy resource consumption, production safety and cost.The calcining heat of rotary kiln is again the important parameter of its running status of reflection, plays conclusive effect to ensureing product quality, realizing the safe and stable operation of rotary kiln and reducing labour intensity.Because rotary kiln body is long and in constantly rotating, the temperature of burning zone is very high again, directly utilize sensor to carry out measuring very difficult, therefore, the temperature how obtaining burning zone exactly becomes a subject matter of rotary kiln process control.

The main estimation adopting manual type to carry out calcining belt temperature both at home and abroad at present, exist estimated accuracy low, there is no scientific basis, easily cause the problems such as security incident, and soft-measuring technique is utilized and easily surveys process variable and predicted the process variable being difficult to directly measure by various calculating.And for the hard measurement problem of calcined by rotary kiln band temperature, current Chinese scholars has carried out a large amount of research work, flexible measurement method based on artificial neural network mainly can be divided into (see ZHANG Y, ZHU J, WANG L M.Temperature prediction model of rotary kiln firing zone based on improved BP neural network [C] // Proc.of 2012 International Conference on Intelligent Systems Design and Engineering Applications, 2012:549-552 and LIU X F, LIU T B, GAO D G.Temperature control in cement rotary kiln with neural network-based heuristic dynamic programming [C] // Advances in Neural Networks-6th International Symposium on Neural Networks, 2009:1078-1086 etc.), based on the flexible measurement method of radiation theory (see LIU J X, ZHU Y L, SHEN Z, et al.Hybrid recognition method for burning zone condition of rotary kiln [J] .Acta Automatica Sinica, 2012, 38 (7): 1153-1161 and Zhang little Gang, Chen Hua, Zhang Jing, Deng. based on the rotary kiln sintered temperature intelligent PREDICTIVE CONTROL [J] of image feedback. control theory and application, 2007, 2 (6): 995-998 etc.) and data-driven method (see Lin B, Recke B, Schmidt T M, et al.Data-driven soft sensor design with multiple-rate sampled data:a comparative study [J] .Industrial & Engineering Chemistry Research, 2009, 48 (12): 5379-5387 and Kaneko H, Arakawa M, Funatsu K.Development of a new soft sensor method using in dependent component analysis and partial least squares [J] .AIChE Journal, 2009, 55 (1): 87-98) etc.But the temperature hard measurement based on artificial neural network is difficult to obtain good prediction effect and unstable, is easily absorbed in local optimum, while too rely on training sample.Based on the temp measuring method of radiation temperature measurement principle according to the heat radiation of testee, under the method is based upon the prerequisite of black matrix hypothesis, and rotary kiln is not strict black matrix, and therefore, the temperature error measuring burning zone is larger.And algorithm and speed of convergence is slow when needing data volume Datong District based on the flexible measurement method of data-driven.

Summary of the invention:

Goal of the invention:

First the present invention utilizes the many kinds of parameters data of KPCA method to rotary kiln collection in worksite to carry out the extraction of pivot composition, by pivot composition, minimum LSSVM is utilized to carry out the foundation of soft-sensing model, utilize the parameter of PSO algorithm to least square method supporting vector machine to be optimized simultaneously, reach method and can obtain satisfied prediction effect and higher precision of prediction, also reduce the modeling optimization time simultaneously, be more suitable for on-the-spot application.

Technical scheme:

A kind of flexible measurement method of calcined by rotary kiln band temperature, it is characterized in that: lime rotary kiln working site is collected data carry out analyzing and processing by core pivot element analysis KPCA, again the data processed are carried out modeling by least square method supporting vector machine LSSVM, checking sample set input least square method supporting vector machine LSSVM model the most at last after core pivot element analysis KPCA process completes the hard measurement of calcining belt temperature, and concrete steps are as follows:

Step 1: the gas flow, temperature of kiln head, kiln end temperature, the primary heater push rod time service parameter that gather rotary kiln working site, these data are reached by Rye the outlier data sample that the change of sensor fault, production environment produces by criterion reject, input data are calculated to the mean value of sample:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}X\left(i\right)$

Then the standard deviation of sample sequence is calculated:

$\mathrm{\σ}=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(X\left(i\right)-\mathrm{\μ})}^2}{N-1}}$

For sample sequence, meet following formula and then think reasonably:

μ-3σ＜X(i)＜μ+3σ

Wherein X (i) is input amendment, and μ is sample average, and σ is standard deviation; The training data not meeting this formula thinks unreasonable data, is rejected, and then adopts the method additional sample data of one dimension interpolation;

Step 2: core pivot element analysis KPCA dimension-reduction treatment is carried out to the data after process, extract according to threshold value E and obtain n major component:

For given rotary kiln running parameter data sample sequence x _k, (k=1,2 ..., m), be mapped to feature space Φ (x _k), Φ () is Nonlinear Mapping, and m is input amendment sequence length, calculates covariance matrix C:

$C=\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}{\left(x_k\right)}^T$

The calculating of major component realizes by separating eigenwert, finds the eigen vector meeting following formula:

λV＝CV

The sample vector linear expression that can be mapped to feature space for V is following formula, wherein α _kfor equation coefficient:

$V=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)$

The matrix K of a definition m × m _i,j:

K _ij＝K(x _i,x _j)＝(Φ(x _i),Φ(x _j)),i,j＝1,2,…,m

Calculate α _kequation coefficient can change the proper vector and eigenwert of asking matrix K into:

mλa＝Ka

Wherein a is α _kthe column vector formed; To V normalization, this time series Φ (x _k) being mapped as on V:

$h_k\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}K(x_k,x)$

Wherein h _kx () is Non-linear Principal Component component;

Ratio represent component h _kpercentage contribution in overall variance; Rotary kiln running parameter major component choose by following formula:

$[\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_k/\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_k]>E$

In formula: n is the major component quantity chosen, E is for choosing percentage threshold, and m is all composition quantity of rotary kiln running parameter sample sequence;

Step 3: the foundation data sample processed being carried out soft-sensing model by least square method supporting vector machine LSSVM, using number of principal components according to the input as least square method supporting vector machine LSSVM, calcined by rotary kiln band temperature is as output, carry out the training of least square method supporting vector machine LSSVM model, utilize particle cluster algorithm PSO algorithm optimization least square method supporting vector machine LSSVM model parameter γ and σ ², detailed process is as follows:

The parameter of initialization particle cluster algorithm: inertia weight, Studying factors, population size, maximum iteration time;

The least square method supporting vector machine LSSVM corresponding to each particle is used to predict training sample respectively, using the fitness value of the root-mean-square error of predicting temperature values and actual temperature value as each particle, again the current fitness value of each particle and the optimal-adaptive angle value of this particle self are compared, if current particle is more excellent, then the position selecting particle current is as the optimal location of this particle; The present invention using the error mean square root of the predicted value of calcining belt temperature and actual value as fitness function, as follows:

$\mathrm{fitness}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y_i})}^2}$

Wherein y _ifor the actual value of calcining belt temperature, for the predicted value of calcining belt temperature;

The optimal location fitness value of each particle self optimal location fitness value and colony is compared, if this particle is more excellent, then using the optimal location of the optimal location of this particle as colony;

Utilize the following formula more speed of new particle and position;

$\begin{array}{c}{v}_{\mathrm{id}}^{k+1}=\mathrm{\ω}\×v_{\mathrm{id}}^k+c_1\×{\mathrm{rand}}_1^k\×\\ ({\mathrm{Pbest}}_{\mathrm{id}}^k-x_{\mathrm{id}}^k)+c_2\×{\mathrm{rand}}_2^k\×({\mathrm{Gbset}}_d^k-x_{\mathrm{id}}^k)\end{array}$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

Check whether the condition meeting optimizing and terminate, that is: reach the maximum iteration time preset or the precision preset, if meet, then optimizing terminates, and obtains optimum solution; Otherwise go to step (2), continue new round search;

Step 4: after least square method supporting vector machine LSSVM model is set up, after checking sample set is carried out core pivot element analysis KPCA process, input least square method supporting vector machine LSSVM model carries out the prediction of calcining belt temperature.

This method is also applicable to the rotary kiln of other type.

Advantageous effect:

The present invention is directed to lime rotary kiln, its working site can collect a large amount of supplemental characteristics, comprises gas flow, push rod time, kiln head and tail temperature etc., but these data are different for the impact of calcining belt temperature.The present invention introduces core pivot element analysis (KPCA) and processes these data, and the key parameter that finding out affects calcining belt temperature carries out modeling, reduces the dimension of sample, reduces the complexity of modeling.Least square method supporting vector machine (LSSVM) structure is simple, have perfect theoretical foundation, only need little sample, is applicable to very much the hard measurement of calcined by rotary kiln band temperature.The present invention simultaneously utilizes particle cluster algorithm (PSO) to be optimized modeling method of least squares support parameter, improves the precision of modeling and prediction.

Accompanying drawing explanation

Fig. 1 is calcined by rotary kiln band temperature flexible measurement method one-piece construction figure of the present invention;

Fig. 2 is calcined by rotary kiln band temperature flexible measurement method data prediction process flow diagram of the present invention;

Fig. 3 is calcined by rotary kiln band temperature flexible measurement method PSO algorithm optimization process flow diagram of the present invention;

Fig. 4 is that calcined by rotary kiln band temperature flexible measurement method PSO of the present invention optimizes LSSVM modeling and prediction process flow diagram.

Embodiment

A kind of flexible measurement method of calcined by rotary kiln band temperature, it is characterized in that: lime rotary kiln working site is collected data carry out analyzing and processing by core pivot element analysis KPCA, again the data processed are carried out modeling by least square method supporting vector machine LSSVM, checking sample set input least square method supporting vector machine LSSVM model the most at last after core pivot element analysis KPCA process completes the hard measurement of calcining belt temperature, and concrete steps are as follows:

Step 1: the gas flow, temperature of kiln head, kiln end temperature, the primary heater push rod time service parameter that gather rotary kiln working site, these data are reached by Rye the outlier data sample that the change of sensor fault, production environment produces by criterion reject, input data are calculated to the mean value of sample:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}X\left(i\right)$

Then the standard deviation of sample sequence is calculated:

$\mathrm{\σ}=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(X\left(i\right)-\mathrm{\μ})}^2}{N-1}}$

For sample sequence, meet following formula and then think reasonably:

μ-3σ＜X(i)＜μ+3σ

Wherein X (i) is input amendment, and μ is sample average, and σ is standard deviation; The training data not meeting this formula thinks unreasonable data, is rejected, and then adopts the method additional sample data of one dimension interpolation;

Step 2: core pivot element analysis KPCA dimension-reduction treatment is carried out to the data after process, extract according to threshold value E and obtain n major component:

For given rotary kiln running parameter data sample sequence x _k, (k=1,2 ..., m), be mapped to feature space Φ (x _k), Φ () is Nonlinear Mapping, and m is input amendment sequence length, calculates covariance matrix C:

$C=\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}{\left(x_k\right)}^T$

The calculating of major component realizes by separating eigenwert, finds the eigen vector meeting following formula:

λV＝CV

The sample vector linear expression that can be mapped to feature space for V is following formula, wherein α _kfor equation coefficient:

$V=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)$

The matrix K of a definition m × m _i,j:

K _ij＝K(x _i,x _j)＝(Φ(x _i),Φ(x _j)),i,j＝1,2,…,m

Calculate α _kequation coefficient can change the proper vector and eigenwert of asking matrix K into:

mλa＝Ka

Wherein a is α _kthe column vector formed; To V normalization, this time series Φ (x _k) being mapped as on V:

$h_k\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}K(x_k,x)$

Wherein h _kx () is Non-linear Principal Component component;

Ratio represent component h _kpercentage contribution in overall variance; Rotary kiln running parameter major component choose by following formula:

$[\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_k/\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_k]>E$

In formula: n is the major component quantity chosen, E is for choosing percentage threshold, and m is all composition quantity of rotary kiln running parameter sample sequence;

Step 3: the foundation data sample processed being carried out soft-sensing model by least square method supporting vector machine LSSVM, using number of principal components according to the input as least square method supporting vector machine LSSVM, calcined by rotary kiln band temperature is as output, carry out the training of least square method supporting vector machine LSSVM model, utilize particle cluster algorithm PSO algorithm optimization least square method supporting vector machine LSSVM model parameter γ and σ ², detailed process is as follows:

The parameter of initialization particle cluster algorithm: inertia weight, Studying factors, population size, maximum iteration time;

The least square method supporting vector machine LSSVM corresponding to each particle is used to predict training sample respectively, using the fitness value of the root-mean-square error of predicting temperature values and actual temperature value as each particle, again the current fitness value of each particle and the optimal-adaptive angle value of this particle self are compared, if current particle is more excellent, then the position selecting particle current is as the optimal location of this particle; The present invention using the error mean square root of the predicted value of calcining belt temperature and actual value as fitness function, as follows:

$\mathrm{fitness}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y_i})}^2}$

Wherein y _ifor the actual value of calcining belt temperature, for the predicted value of calcining belt temperature;

The optimal location fitness value of each particle self optimal location fitness value and colony is compared, if this particle is more excellent, then using the optimal location of the optimal location of this particle as colony;

Utilize the following formula more speed of new particle and position;

$\begin{array}{c}{v}_{\mathrm{id}}^{k+1}=\mathrm{\ω}\×v_{\mathrm{id}}^k+c_1\×{\mathrm{rand}}_1^k\×\\ ({\mathrm{Pbest}}_{\mathrm{id}}^k-x_{\mathrm{id}}^k)+c_2\×{\mathrm{rand}}_2^k\×({\mathrm{Gbset}}_d^k-x_{\mathrm{id}}^k)\end{array}$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

Check whether the condition meeting optimizing and terminate, that is: reach the maximum iteration time preset or the precision preset, if meet, then optimizing terminates, and obtains optimum solution; Otherwise go to step (2), continue new round search;

Step 4: after least square method supporting vector machine LSSVM model is set up, after checking sample set is carried out core pivot element analysis KPCA process, input least square method supporting vector machine LSSVM model carries out the prediction of calcining belt temperature.

Below in conjunction with accompanying drawing, the present invention is described in further detail.

As shown in the figure, Fig. 1 is calcined by rotary kiln band temperature flexible measurement method one-piece construction figure of the present invention; Fig. 2 is calcined by rotary kiln band temperature flexible measurement method data prediction process flow diagram of the present invention.

Step 1: gather the foundation that rotary kiln working site data carry out (comprising primary heater push rod time, kiln end temperature etc.) soft-sensing model.The change of sensor fault, production environment may be there is in rotary kiln working site, this can make to there are outlier data in the sample of collection, these data can reduce the accuracy of soft-sensing model modeling, therefore these outlier data are needed to reject, the present invention adopts Rye to reach criterion and removes irrational data, input data is calculated to the mean value of sample:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}X\left(i\right)$

Then the standard deviation of sample sequence is calculated

$\mathrm{\σ}=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(X\left(i\right)-\mathrm{\μ})}^2}{N-1}}$

For sample sequence, meet following formula and then think reasonably

μ-3σ＜X(i)＜μ+3σ

Wherein X (i) is input amendment, and μ is sample average, and σ is standard deviation.The training data not meeting this formula thinks unreasonable data, is rejected, and then adopts the method additional sample data of one dimension interpolation;

Step 2: carry out KPCA dimension-reduction treatment to the data after process, extracts according to threshold value E and obtains n major component, can be specifically described as follows

For given rotary kiln running parameter data sample sequence x _k, (k=1,2 ..., m), be mapped to feature space Φ (x _k), Φ () is Nonlinear Mapping, and m is input amendment sequence length, calculates covariance matrix C

$C=\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}{\left(x_k\right)}^T$

The calculating of major component realizes by separating eigenwert, finds the eigen vector meeting following formula

λV＝CV

The sample vector linear expression that can be mapped to feature space for V is following formula, wherein α _kfor equation coefficient

$V=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)$

The matrix K of a definition m × m _i,j

K _ij＝K(x _i,x _j)＝(Φ(x _i),Φ(x _j)),i,j＝1,2,…,m

Calculate α _kequation coefficient can change the proper vector and eigenwert of asking matrix K into

mλa＝Ka

Wherein a is α _kthe column vector formed.To V normalization, this time series Φ (x _k) being mapped as on V

$h_k\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}K(x_k,x)$

Wherein h _kx () is Non-linear Principal Component component.

Ratio represent component h _kpercentage contribution in overall variance.Choosing by following formula of rotary kiln running parameter major component

$[\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_k/\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_k]>E$

In formula: n is the major component quantity chosen, E is for choosing percentage threshold, and m is all composition quantity of rotary kiln running parameter sample sequence.

Step 3: using number of principal components according to the input as LSSVM, calcined by rotary kiln band temperature, as output, carries out the training of LSSVM model, utilizes PSO algorithm optimization LSSVM model parameter γ and σ ², as shown in Figure 3, detailed process is as follows:

The parameter of initialization particle cluster algorithm, comprising: inertia weight, Studying factors, population size, maximum iteration time etc.;

The LSSVM corresponding to each particle is used to predict training sample respectively, using the fitness value of the root-mean-square error of predicting temperature values and actual temperature value as each particle, the more current fitness value of each particle and the optimal-adaptive angle value of this particle self are compared.If current particle is more excellent, then the position selecting particle current is as the optimal location of this particle; The present invention using the error mean square root of the predicted value of calcining belt temperature and actual value as fitness function, as follows:

$\mathrm{fitness}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y_i})}^2}$

Wherein y _ifor the actual value of calcining belt temperature, for the predicted value of calcining belt temperature.

The optimal location fitness value of each particle self optimal location fitness value and colony is compared, if this particle is more excellent, then using the optimal location of the optimal location of this particle as colony;

Utilize the following formula more speed of new particle and position:

$\begin{array}{c}{v}_{\mathrm{id}}^{k+1}=\mathrm{\ω}\×v_{\mathrm{id}}^k+c_1\×{\mathrm{rand}}_1^k\×\\ ({\mathrm{Pbest}}_{\mathrm{id}}^k-x_{\mathrm{id}}^k)+c_2\×{\mathrm{rand}}_2^k\×({\mathrm{Gbset}}_d^k-x_{\mathrm{id}}^k)\end{array}$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

Check whether the condition (reaching the maximum iteration time preset or the precision preset) meeting optimizing and terminate.If meet, then optimizing terminates, and obtains optimum solution; Otherwise go to step (2), continue new round search.

After step 4:LSSVM model is set up, after checking sample set is carried out KPCA process, input LSSVM model carries out the prediction of calcining belt temperature, as shown in Figure 4.

Claims (2)

1. the flexible measurement method of a calcined by rotary kiln band temperature, it is characterized in that: lime rotary kiln working site is collected data carry out analyzing and processing by core pivot element analysis KPCA, again the data processed are carried out modeling by least square method supporting vector machine LSSVM, checking sample set input least square method supporting vector machine LSSVM model the most at last after core pivot element analysis KPCA process completes the hard measurement of calcining belt temperature, and concrete steps are as follows:

Step 1: the gas flow, temperature of kiln head, kiln end temperature, the primary heater push rod time service parameter that gather rotary kiln working site, these data are reached by Rye the outlier data sample that the change of sensor fault, production environment produces by criterion reject, input data are calculated to the mean value of sample:

$\mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}X\left(i\right)$

Then the standard deviation of sample sequence is calculated:

$\mathrm{\σ}=\sqrt{\frac{\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(X\left(i\right)-\mathrm{\μ})}^2}{N-1}}$

For sample sequence, meet following formula and then think reasonably:

μ-3σ＜X(i)＜μ+3σ

Wherein X (i) is input amendment, and μ is sample average, and σ is standard deviation; The training data not meeting this formula thinks unreasonable data, is rejected, and then adopts the method additional sample data of one dimension interpolation;

Step 2: core pivot element analysis KPCA dimension-reduction treatment is carried out to the data after process, extract according to threshold value E and obtain n major component:

For given rotary kiln running parameter data sample sequence x _k, (k=1,2 ..., m), be mapped to feature space Φ (x _k), Φ () is Nonlinear Mapping, and m is input amendment sequence length, calculates covariance matrix C:

$C=\frac{1}{m}\underset{k=1}{\overset{m}{\mathrm{\Σ}}}\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}{\left(x_k\right)}^T$

The calculating of major component realizes by separating eigenwert, finds the eigen vector meeting following formula:

λV＝CV

The sample vector linear expression that can be mapped to feature space for V is following formula, wherein α _kfor equation coefficient:

$V=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)$

The matrix K of a definition m × m _i,j:

$K_{\mathrm{ij}}=K(x_i,x_j)=(\mathrm{\Φ}\left(x_i\right),\mathrm{\Φ}\left(x_j\right)),i,j=\mathrm{1,2},...,m$

Calculate α _kequation coefficient can change the proper vector and eigenwert of asking matrix K into:

mλa＝Ka

Wherein a is α _kthe column vector formed; To V normalization, this time series Φ (x _k) being mapped as on V:

$h_k\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_k\mathrm{\Φ}\left(x_k\right)\mathrm{\Φ}\left(x\right)=\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}K(x_k,x)$

Wherein h _kx () is Non-linear Principal Component component;

Ratio represent component h _kpercentage contribution in overall variance; Rotary kiln running parameter major component choose by following formula:

$[\underset{k=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\λ}}_k/\underset{k=1}{\overset{m}{\mathrm{\Σ}}}{\mathrm{\λ}}_k]>E$

In formula: n is the major component quantity chosen, E is for choosing percentage threshold, and m is all composition quantity of rotary kiln running parameter sample sequence;

Step 3: the foundation data sample processed being carried out soft-sensing model by least square method supporting vector machine LSSVM, using number of principal components according to the input as least square method supporting vector machine LSSVM, calcined by rotary kiln band temperature is as output, carry out the training of least square method supporting vector machine LSSVM model, utilize particle cluster algorithm PSO algorithm optimization least square method supporting vector machine LSSVM model parameter γ and σ ², detailed process is as follows:

The parameter of initialization particle cluster algorithm: inertia weight, Studying factors, population size, maximum iteration time;

The least square method supporting vector machine LSSVM corresponding to each particle is used to predict training sample respectively, using the fitness value of the root-mean-square error of predicting temperature values and actual temperature value as each particle, again the current fitness value of each particle and the optimal-adaptive angle value of this particle self are compared, if current particle is more excellent, then the position selecting particle current is as the optimal location of this particle; The present invention using the error mean square root of the predicted value of calcining belt temperature and actual value as fitness function, as follows:

$\mathrm{fitness}=\sqrt{\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(y_i-\stackrel{\‾}{y_i})}^2}$

Wherein y _ifor the actual value of calcining belt temperature, for the predicted value of calcining belt temperature;

The optimal location fitness value of each particle self optimal location fitness value and colony is compared, if this particle is more excellent, then using the optimal location of the optimal location of this particle as colony;

Utilize the following formula more speed of new particle and position;

$\begin{array}{c}{v}_{\mathrm{id}}^{k+1}=\mathrm{\ω}\×v_{\mathrm{id}}^k+c_1\×\mathrm{ran}{d}_1^k\×\\ ({\mathrm{Pbest}}_{\mathrm{id}}^k-x_{\mathrm{id}}^k)+c_2\×{\mathrm{rand}}_2^k\×({\mathrm{Gbest}}_d^k-x_{\mathrm{id}}^k)\end{array}$

$x_{\mathrm{id}}^{k+1}=x_{\mathrm{id}}^k+v_{\mathrm{id}}^{k+1}$

Check whether the condition meeting optimizing and terminate, that is: reach the maximum iteration time preset or the precision preset, if meet, then optimizing terminates, and obtains optimum solution; Otherwise go to step (2), continue new round search;

Step 4: after least square method supporting vector machine LSSVM model is set up, after checking sample set is carried out core pivot element analysis KPCA process, input least square method supporting vector machine LSSVM model carries out the prediction of calcining belt temperature.

2. the flexible measurement method of a kind of calcined by rotary kiln band temperature according to claim 1, is characterized in that: this method is also applicable to the rotary kiln of other type.