CN101038277B - Soft measurement method for coal power fineness in powdering producer - Google Patents

Abstract

A pulverized coal fineness soft-sensing method in a powder-making process based on a least squares support vector machine depends on a hardware platform, measuring instruments and a computer system used for performing a software calculation, wherein said software performs a soft-sensing through real-time process data obtained by a controlling computer of a data collector. Said method comprises the steps of selecting assistant variables, determining the structure of sample data, determining sample components, building a soft-sensing model based on the least squares support vector machine and performing the model training, parameter selection as well as the model validation and the model parameter emendation. The present invention is applied simply and easily, has a lower cost, has more accurate soft-sensing results, and helps to achieve an optimal control and an optimal operation of a pulverizing process.

Description

Pulverizing process fineness of pulverized coal flexible measurement method based on least square-support vector machine

Technical field

The invention belongs to the automatic measurement technical field, particularly to the storage low speed coal mill pulverized coal preparation system of forming by low speed coal mill, mill separator and cloth bag gatherer that is used in the pulverizing process that coal dust grinds final products---the fineness index of coal dust is carried out the method for soft measurement.

Background technology

During powder process was produced, the middle storage low speed coal mill of being made up of low speed coal mill, mill separator and cloth bag gatherer was widely used in coal dust is ground in the fineness scope of technological requirement.The fineness of coal dust is the important technical index of examination pulverizing process, is to improve the quality of products and the key factor of economic benefit, occupies crucial status for process control such as the generating of adopting coal-fired mode, sintering.At present, in the commercial Application of the online detection of fineness of pulverized coal, because the restriction of physical condition, do not carry out type approval test, do not have actual operability, do not have ripe pick-up unit, the main mode off-line of sampling, chemical examination that relies on obtains in actual production process.The deficiency of this method is: 1. the human factor influence is big during manually-operated, and the objectivity of measurement result is poor; The time interval of Ce Lianging long, the time of measurement feedback is also long, the information that therefore obtains lacks directive significance to operating personnel; 3. the coal dust leakage phenomenon unavoidably occurs when coal dust being taken a sample, cause environmental pollution.Because common neural net method is adopted in sample data limited amount and the bigger influence of mushing error to training, the speed of convergence of model is slow, and occurs the over-fitting phenomenon easily.

Summary of the invention

In order to solve the deficiency of existing pulverizing process fineness of pulverized coal measuring method, the invention provides a kind of fineness of pulverized coal flexible measurement method based on least square-support vector machine, the measurement parameter of the auxiliary variable that provides by conventional on-line measurement instrument, provide the estimated value of current fineness of pulverized coal, moving with optimization for the Optimizing operation of pulverizing process provides the critical process index.

The inventive method is made up of hardware platform and Survey Software, and wherein the hardware platform core is made up of low speed coal mill, mill separator, cloth bag gatherer and relevant device, the computer system that has been equipped with measurement instrument simultaneously and has carried out computed in software.The connection of its hardware is that input end and feeder, hot blast and the cold air duct of low speed coal mill joins, the low speed coal mill output terminal is connected with mill separator, the powder retrieving pipe road of mill separator is connected with the coal pulverizer output terminal, outlet is joined with sack cleaner, coal dust realizes that through the cloth bag gatherer wind powder separates, and coal dust is sent into Pulverized Coal Bin by conveying device.Detailed structure as shown in Figure 1.With the pulverized coal preparation system that low speed coal mill, mill separator, cloth bag gatherer are formed, its measurement instrument comprises:

A thermopair is installed on hot air duct, is used for on-line measurement hot blast temperature T _RFWD

A thermopair is installed on the coal pulverizer outlet conduit, is used for on-line measurement coal pulverizer outlet coal dust temperature T _CKWD

On the feed pipeline of coal pulverizer and cloth bag gatherer, discharge pipe, four pressure gauges are installed respectively, are used for on-line measurement coal pulverizer inlet and outlet pressure P _MRPL, P _MCYLWith cloth bag gatherer inlet and outlet pressure P _SRYL, P _SCYLBe used to calculate coal pulverizer and sack cleaner differential pressure P _MJCY, P _SJQCY

The feedback signal of feeder rotating speed is used to estimate coal-supplying amount W _GMJZSIf the measuring-signal of coal-supplying amount is arranged, directly use the detected value of coal-supplying amount.Use feeder rotating speed W in this instructions _GMJZSEstimate coal-supplying amount.

This grinding system disposed control computer (distributed computer control system (DCS) or programmable logic controller (PLC) (PLC) or industrial control computer (IPC)) or data acquisition unit simultaneously, in order to gather the signal of measurement instrument.

Soft Survey Software of the present invention both may operate on the programmable logic controller (PLC), on the supervisory control comuter of computer control system, also can run on independently on the computing machine, this software is by carrying out communication with control computer or data acquisition unit, obtain real-time process data, and provide the estimated result of fineness of pulverized coal.

The inventive method may further comprise the steps: the selection of (1) auxiliary variable, and the structure of (2) sample data, the composition of (3) sample, (4) based on least square-support vector machine soft-sensing model, the correction of (5) model parameter.

(1) selection of auxiliary variable, from the experience that influences fineness of pulverized coal and the correlative factor analysis of theoretical analysis fineness of pulverized coal, the present invention selects auxiliary variable to obtain by the online detection of instrument, comprises following auxiliary variable:

Hot blast temperature T _RFWD

Coal pulverizer outlet temperature T _CKWD

Coal pulverizer differential pressure P _MJCY: coal pulverizer inlet pressure P _MRPLWith coal pulverizer top hole pressure P _MCYLDifference;

The differential pressure maximal value P of gatherer work period _MAXSJQCY: cloth bag gatherer inlet pressure P _SRYLWith cloth bag gatherer outlet power P _SCYLThe maximum difference of work period;

Feeder rotating speed W _GMJZS, or coal-supplying amount.

(2) structure of sample data:

The coal powder grinding process is formed sample according to following structure, and collects sample data.Sample is expressed as { x _k, y _k, x wherein _kBe the input of sample, the auxiliary variable of both having chosen---hot blast temperature T _RFWD, coal pulverizer outlet temperature T _CKWD, coal pulverizer differential pressure P _MJCY, the gatherer work period differential pressure maximal value P _MAXSJQCY, feeder rotating speed W _GMJZSSample is output as leading variable to be estimated---fineness of pulverized coal L _MFXDSample collection interrecord structure such as table 1, the time is the time that sample obtains, and for reducing Error and Assay, adopts three results of laboratory test to carry out sample according to same sample fineness of pulverized coal and accepts or rejects, average at last:

Table 1 sample data structure

Consider that sample data should be representative, and coverage broad as far as possible, at least should comprise the industrial object normal range of operation, by manual adjustments valve area, feeder frequency, in producing the scope that allows, change the working point of system as far as possible, sample examination after each operating conditions change system operates steadily.

(3) composition of sample:

The present invention is used for the sample data of soft-sensing model and forms by three groups: model training sample, error training sample and checking sample are formed.Consider that the training sample data should be representative, and coverage broad as far as possible, at least should comprise the industrial object normal range of operation, by manual adjustments exhaust valve aperture, feeder frequency, in producing the scope that allows, change the working point of system, after each operating conditions changes, sample examination after system operates steadily.In actual production process, to fineness of pulverized coal when chemical examination selective examination according to table one record, obtain the sample data of some, this partial data is divided into two groups, one group of parameter that the training that is used for the model error minimum comes preference pattern, another group is used for verification of model.

(4) based on the least square method supporting vector machine soft-sensing model

Of the present invention based on the least square method supporting vector machine soft-sensing model, the basic procedure block diagram of its model training as shown in Figure 2.Comprise following detailed step:

(A) initialization: carry out the initialization of all variablees.

(B) carry out soft measurement? if then go to (C), according to input data and soft-sensing model forecasting coal powder fineness; If not, then go to (G), carry out the process of model training, checking and parameter correction.

Step (C) to (F) is the soft-sensing model application flow, according to input data and soft-sensing model forecasting coal powder fineness.

(C) read the current working data;

The just online characterising parameter that reads the current working auxiliary variable that to carry out the soft measurement of fineness.

(D) input data normalization:

The input variable supplemental characteristic is carried out standardization.

(E) soft-sensing model calculates

Treated input data are updated to the soft measurement result of the following Model Calculation fineness of pulverized coal after process is trained, verified.

$y\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\Σ}}}{\mathrm{\α}}_{k}k(x,x_k)+b$

X in the formula _k---the input of sample, the auxiliary variable of both having chosen---hot blast temperature T _RFWD, coal pulverizer outlet temperature T _CKWD, coal pulverizer differential pressure P _MJCY, the gatherer work period differential pressure maximal value P _MAXSJQCY, feeder rotating speed W _GMJZS

Y (x)---leading variable to be estimated---fineness of pulverized coal L _MFXD

α _k, b---the model coefficient after the training;

k＝1，2，...，n；

(F) demonstration and the soft measurement result of preservation fineness

The soft measured value of the fineness of pulverized coal that least square-supporting vector machine model is calculated is presented on the man-machine interface, and is saved in the historical data base with current other parametric descriptions, relatively reaches other for the reality chemical examination and manipulates.

In order to give full play to the function of fineness soft-sensing model of the present invention, soft process of measurement should be able to read current working in real time automatically and describe, concrete time limit length is determined according to concrete process condition, automatically the current working that reads is described then and carried out the soft measurement calculating of fineness, soft measurement result is presented on the picture that operator or tradesman see easily (can be " optimum management " picture of monitored picture), and result and other data are saved in the Relational database.

Step (G) to (T) is the process of model training, checking and parameter correction.

(G) preference pattern error punishment parameter set and nuclear parameter collection

Determine that as the case may be this interval should comprise the optimized parameter of model, so that therefrom select the best model parameter to the error punishment parameter of model training use and the interval range of nuclear parameter.

(H) all sample data standardization

Training sample, error assessment sample and modelling verification sample are carried out standardization, be used for the input of model training, error assessment and modelling verification.

(I) read in the training sample data

Read in the sample data of preparing to be used for model training

(J) adjustment model error punishment parameter and nuclear parameter

From the lower limit of the scope setting of error punishment parameter and nuclear parameter, each parameter circulation increases a step-length at every turn, as adjusted parameter, is used to set up corresponding model this group parameter is carried out error assessment.

(K) set up model

Modelling process prescription based on least square-support vector machine is as follows:

The present invention is that example is set up soft-sensing model to adopt radially basic kernel function (RBF), and this kernel function form is:

$k(x_k,x_l)=\exp (-\frac{{\left|\right|x_k-x_l\left|\right|}^2}{2{\mathrm{\δ}}^2})$

In the formula: δ---nuclear parameter;

This soft-sensing model has according to following derivation: for given training set is { x _k, y _k} _K=1 ⁿ, x wherein _k∈ R ^dy _k∈ R, φ (.), d are the auxiliary variable number.The basic idea about modeling of support vector machine is by Nonlinear Mapping φ (.), importing sample from input space R ^dBe mapped to feature space φ (x)=(φ ₁(x), φ ₂(x) ... φ _n(x)), thus the nonlinear estimation problem is converted into the linear function estimation problem of high-dimensional feature space.Promptly adopt the function of following form that pattern function is estimated:

f(x)＝W ^Tφ(x _k)+b，(W∈R ^dn，b∈R)

In the formula: W ^T---weight vector

B---amount of bias

According to problem solving target and structuring risk minimum is principle, and following formula need satisfy following formula simultaneously:

|y _k-W ^Tφ(x _k)+b|≤ε， $\min J=\frac{1}{2}{W}^{T}W$

Adopt least square-support vector machine method, promptly defining the error loss function is the quadratic term e of error _k ², the problems referred to above can be expressed as in weight space W, $\min J(W,e)=\frac{1}{2}{W}^{T}W+\frac{1}{2}\mathrm{\&gamma;}\underset{i=1}{\overset{\mathrm{<figure-callout\; id="2"\; label="n\; e\; k"\; filenames="S07111021320070514E000011.png"\; state="\{\{state\}\}">n</figure-callout>}}{\mathrm{\&Sigma;}}}{e}_k^{}{}_2^{}$ Satisfy constraint condition: y _k=W ^T

(x _k)+b+e _k(k=1,2 ..., n).Wherein γ is the error penalty factor.Find the solution above-mentioned optimization problem with Lagrangian method, the definition Lagrangian function is as follows:

α wherein _k(k=1,2 ... n) be Lagrange multiplier.The problems referred to above are summed up as quadratic programming problem, order

$\frac{\&PartialD;L}{\&PartialD;e_k}=0,$ $\frac{\&PartialD;L}{\&PartialD;W}=0,$ $\frac{\&PartialD;L}{\&PartialD;{\mathrm{\&alpha;}}_k}=0,$ $\frac{\&PartialD;L}{\&PartialD;b}=0$

Eliminate W, e has $\left[\begin{array}{cc}0& l^T\\ l& \mathrm{\&Omega;}+\frac{1}{\mathrm{\&gamma;}}I\end{array}\right]\left\{\begin{array}{c}b\\ a\end{array}\right\}=\left\{\begin{array}{c}0\\ y\end{array}\right\},$ Wherein: y=[y ₁, y ₂, y ₃... y _n]; L=[1,1,1 ... 1];

α＝[α ₁，α ₂，α ₃，...α _n]；Ω _kl＝ (x _k) ^T

(x _l)；k＝1，2，...，n。

Determine coefficient b and α by following formula, obtain soft-sensing model and be:

$y\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x,x_k)+b$

(L) read the error assessment sample data

Read in one group of sample data preparing to be used for error assessment

(M) recording error evaluation result and parameter

Get error training sample data collection S, comprising number of samples is l, the definition error function:

$e1=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e_i}^2=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{(y_i-(\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x_i,x_k)+b))}^2$

$e2=\max {(y_i-(\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x_i,x_k)+b))}^2$

I=1 wherein, 2 ..., l.Select the final error evaluation function to be:

e(γ，δ)＝min(e1+ηe2)

In the formula: γ---error punishment parameter;

δ---nuclear parameter;

η---weight parameter;

Rule of thumb select the weight of mean square deviation and maximum variance, generally can select η=1.The soft-sensing model that utilizes (K) to set up between given parameter region obtains the error assessment index, and the record corresponding parameters.

(N) whether parameter adjustment has arrived the upper limit

If γ+l＞γ _Up(γ is an error punishment parameter, and l is for adjusting step-length, γ _UpFor punishing the upper limit of parameter area) and δ+l＞γ _Up(δ is that nuclear parameter, l are for adjusting step-length, δ _UpThe upper limit for the nuclear parameter scope) satisfies simultaneously, illustrate that then the error assessment work of all parameter combinations is finished; Otherwise repeat the work of (J)～(N).

(O) Select Error is estimated best model parameter

With the error assessment index of record in (M), seek minimal value wherein, be the best model parameter of error assessment, the selection corresponding parameters is a model parameter.

(P) determine model

According to the model parameter of selecting in (O), determine the model training result, determine soft-sensing model.

(Q) read in the checking sample data

Read in one group of sample data preparing to be used for modelling verification

(R) modelling verification

Model in the employing (P) is according to e1 index in (M) and e2 index.

(S) whether the checking precision is satisfied with

Soft-sensing model is verified the output of softer measurement model and actual laboratory values with the checking sample data.If the relative error of model output valve and actual laboratory values is in the error allowed band, this model can be used for actual fineness of pulverized coal On-line Estimation; If the relative error of this model output valve and actual laboratory values is bigger, the analyzing and training sample data continues experiment, suitably increases the training sample data, repeats the work of (B)～(S).

(T) re-construct training sample

The checking precision can not satisfy the needs of soft measurement, and needing increases test number (TN), re-constructs training sample.

(U) finish

(5) correction of model parameter:

As time goes on, the sample data that chemical examination obtains constantly increases, and is further correction model parameter, and improving model accuracy provides possibility.Select typical sample to add training sample to,, then replace original model, otherwise do not change master mould with new model if Xun Lian soft-sensing model can obviously improve the precision of model again.If it is identical or close that new sample data repeats the sample initial conditions, and sample output differs bigger situation, replaces former sample data with new samples, and training pattern is replaced original model again.

The invention has the advantages that: utilize department of computer science to unify online process data that conventional measuring instrument provides, only by a spot of artificial sample, realized the soft measurement based on least square-support vector machine of fineness of pulverized coal.Solved a difficult problem that does not have fineness of pulverized coal on-line detector table to be difficult to online detection; Compare with the hand sampling chemical examination, reduced operating personnel's workload, reduced the uncertainty of the measurement of manual operation introducing, improved measure ageing, the problem of the leakage of pulverized coal that having reduced takes a sample brings.Fuzzyly compare with the neural network flexible measurement method that flexible measurement method has overcome input dimension height among the present invention, sample size is few and with traditional owing to system interference factor multisample is introduced the model generalization that bigger error causes and the problem of over-fitting sometimes.In addition, because the present invention has considered to influence the factor of fineness of pulverized coal comprehensively, adopt the existing conventional detection signal to realize the On-line Estimation of fineness of pulverized coal in a large number, application is simple, easy, and cost is lower, and soft measurement result is also more accurate.This method helps to realize the optimal control of pulverizing process and optimizes operation.

Description of drawings

The flow process of Fig. 1 pulverizing process, measurement instrument and computer configuration figure;

Fig. 2 sets up process flow diagram based on the pulverizing process fineness of pulverized coal soft-sensing model of least square-support vector machine

Among Fig. 1: 1 low speed coal mill, 2 cloth bag gatherers, 3 mill separators, 4 feeders, 5 cloth bag gatherer head pressure gages, 6 cloth bag gatherer inlet/outlet pressure gauges, 7 feeder frequency conversion rotating speeds or coal-supplying amount are weighed, 8 thermopairs, 9 coal pulverizer head pressure gages, 10 coal pulverizer top hole pressure meters, 11 thermal resistances, 12PLC, control computer or special purpose computer

Used label symbol is as follows among Fig. 1:

Hot blast temperature---T _RFWD

Coal pulverizer outlet temperature---T _CKWD

Coal pulverizer inlet pressure---P _MRPL

Coal pulverizer top hole pressure---P _MCYL

Cloth bag gatherer inlet pressure---P _SRYL

Cloth bag gatherer top hole pressure---P _SCYL

Feeder rotating speed or coal-supplying amount---W _GMJZS

In Fig. 1, solid arrow is represented logistics (coal dust, hot blast and cold wind), and dotted line is represented signal flow.

Embodiment

Burning till the used ball mill pulverizing system of rotary kiln with aluminium oxide is example, and the coal powder grinding process is made up of feeder, coal pulverizer, mill separator and equipment such as sack cleaner, Pulverized Coal Bin as shown in Figure 1.Raw coal and hot blast grind in low speed coal mill and are dry, coal dust enters mill separator under the effect of exhaust blower, return powder and turn back to coal pulverizer continuation grinding from the section storehouse, qualified coal dust realizes that through sack cleaner the wind powder separates, and coal dust is sent into Pulverized Coal Bin by conveying device.

The coal pulverizer specification is Φ 3.2 * 9.5m, and designed productive capacity 28t/h, coal pulverizer are divided into oven dry storehouse, ball storehouse, Duan Cangsan storehouse; Main motor model YR80-8/1180, rated power 800KW.

This coal pulverizer system has the requirement of this instructions following measurement instrument is installed, and comprising:

Hot blast temperature;

The coal pulverizer outlet temperature;

The coal pulverizer inlet and outlet pressure;

The gatherer inlet and outlet pressure;

The feeder speed feedback.

Realize the automatic control of basic control loop with Programmable Logic Controller (PLC), and according to the filtering of model needs, calculate:

Hot blast temperature T _RFWD

Coal pulverizer outlet temperature T _CKWD

Coal pulverizer differential pressure P _MJCY

The differential pressure maximal value P of gatherer work period _MAXSJQCY

Feeder rotating speed W _GMJZS

Read above-mentioned process data.WINCC software with Siemens in supervisory control comuter is realized monitoring human-computer interface.

The fineness of pulverized coal of this pulverized coal preparation system (on the 175# sieve) normal range of operation, target setting is 15%～20% when two kilns of supply.

Soft-sensing model realizes that with the modularization programming of the step7 of Siemens PLC C data storage is in the DB piece.At the supply a model modification interface of data of WINCC supervisory system, be used for off-line analysis change model parameter.Soft Survey Software is moved on PLC, has effectively guaranteed the ageing of model output, has made things convenient for the process monitoring of system.

Select auxiliary variable according to the described implementation method of this instructions, obtain experiment sample, set up soft-sensing model, and, obtain model parameter, obtain γ=0.5 and δ=0.8 by the minimum training of error.Under this parameter, calculate a, the value of b is respectively:

b＝19.4587；

a＝[-3.16?-1.11?-1.86?-2.6?0.77?2.6?1.78?17.35?-3.04?5.41?-3.97?-1.95-9.04?0.68?-2.54?1.61?-0.32?-0.97?0.36]

This example is owing to the work period of gatherer is 336s, obtain the maximum differential pressure value of a sub collector every 336s, thereby move once soft process of measurement, carry out once the soft measurement of fineness under should the time operating mode, and soft measurement result is outputed to the system management picture by supervisory control comuter, for the operator or the correlation technique workman carries out parameter optimization and monitor operation provides a good reference, directive function, simultaneously soft measurement result is saved in system's real-time data base.

The present invention promptly based on the fineness of pulverized coal soft-sensing model of least square-support vector machine at the pulverized coal preparation system normal operation period, can estimate coal powder size according to the real time data of process, relative error is no more than 2%, become one have very high practical value, fineness measurement means cheaply.

Claims (5)

1. pulverizing process fineness of pulverized coal index flexible measurement method based on least square-support vector machine, this method relies on hardware platform and measurement instrument and computer system, it is characterized in that obtaining real-time process data by control computer or data acquisition unit, carry out soft measurement, this method may further comprise the steps:

(1) selection of auxiliary variable, the auxiliary variable of selection comprises:

Hot blast temperature T _RFWD

Coal pulverizer outlet temperature T _CKWD

Coal pulverizer differential pressure P _MJCY: coal pulverizer inlet pressure P _MRPLWith coal pulverizer top hole pressure P _MCYLDifference;

The differential pressure maximal value P of gatherer work period _MAXSJQCY: cloth bag gatherer inlet pressure P _SRYLWith cloth bag gatherer top hole pressure P _SCYLThe maximum difference of work period; Feeder rotating speed W _GMJZS, or coal-supplying amount;

(2) determine the structure of sample data, as following table:

X wherein _kBe the input of sample, the auxiliary variable of both having chosen---hot blast temperature T _RFWD, coal pulverizer outlet temperature T _CKWD, coal pulverizer differential pressure P _MJCY, the gatherer work period differential pressure maximal value P _MAXSJQCY, feeder rotating speed W _GMJZS, the output y of sample _kBe leading variable to be estimated---fineness of pulverized coal L _MFXD

(3) composition of sample

The sample data that is used for soft-sensing model is formed by three groups: model training sample, error training sample and checking sample are formed; The model training sample should comprise the industrial object normal range of operation at least, by manual adjustments exhaust valve aperture, feeder frequency, in producing the scope that allows, change the working point of system, after each operating conditions changes, the sample data of sample examination after system operates steadily; To fineness of pulverized coal when chemical examination selective examination according to the record of sample data structure, obtain the sample data of some, this partial data is divided into two groups, one group of parameter that the training that is used for the model error minimum comes preference pattern, another group is used for verification of model;

(4) set up based on least square-support vector machine soft-sensing model, as shown in the formula:

$y\left(x\right)=\underset{k=1}{\overset{k}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x,x_k)+b$

α _k, b---the model coefficient after the training;

k＝1，2，...，n；

Carrying out training, the parameter of model then selects and modelling verification;

(5) correction of model parameter;

Select typical sample to add training sample to,, then replace original model, otherwise do not change master mould with new model if Xun Lian soft-sensing model can obviously improve the precision of model again; If it is identical or close that new sample data repeats the sample initial conditions, and sample output differs bigger situation, replaces former sample data with new samples, and training pattern is replaced original model again.

2. a kind of pulverizing process fineness of pulverized coal index flexible measurement method based on least square-support vector machine according to claim 1 is characterized in that the described modelling of step (4), training, parameter are selected and modelling verification is carried out according to the following steps:

(A) initialization: carry out the initialization of all variablees; (B) carry out soft measurement? if then go to (C), according to input data and soft-sensing model forecasting coal powder fineness; If not, then go to (G), carry out the process of model training and parameter correction; (C) read the current working data; (D) the input data normalization is handled; (E) soft-sensing model calculates; (F) demonstration and the soft measurement result of preservation fineness are carried out (U); (G) preference pattern error punishment parameter set and nuclear parameter collection; (H) all sample data standardization; (I) read in the training sample data; (J) adjustment model error punishment parameter and nuclear parameter; (K) set up model; (L) read the error assessment sample data; (M) recording error evaluation result and parameter; (N) whether parameter adjustment has arrived the upper limit; If parameter adjustment to the upper limit, is selected parameter; If less than the upper limit, adjust the evaluation procedure of new argument repetitive error according to adjusting step-length, return (J); (O) Select Error is estimated best model parameter; (P) determine model; (Q) read in the checking sample data; (R) modelling verification; (S) whether the checking precision is satisfied with, if satisfied, model parameter is selected to finish; If dissatisfied, increase analysis data, re-construct training sample, carry out (T); (T) re-construct training sample; (U) finish.

3. a kind of pulverizing process fineness of pulverized coal index flexible measurement method based on least square-support vector machine according to claim 2, use when it is characterized in that described recording error evaluation result of step (M) and parameter with minor function:

e(γ，δ)＝min(e1+ηe2)

$e1=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e_i}^2=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{(y_i-(\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x_i,x_k)+b))}^2$

$e2=\max {(y_i-(\underset{k=1}{\overset{n}{\mathrm{\&Sigma;}}}{\mathrm{\&alpha;}}_{k}k(x_i,x_k)+b))}^2$

In the formula: γ---error punishment parameter;

δ---nuclear parameter;

η---weight parameter, η=1;

i＝1，2，...，l。

4. a kind of pulverizing process fineness of pulverized coal index flexible measurement method based on least square-support vector machine according to claim 2 is characterized in that step (N) is described and judges whether parameter adjustment should meet following condition: γ+l＞γ in limited time to last _Up, in the formula: γ is that error punishment parameter, l are for adjusting step-length, γ _UpThe upper limit for the punishment parameter area; And δ+l＞γ _Up, in the formula: δ is that nuclear parameter, l are for adjusting step-length, δ _UpBe the upper limit of nuclear parameter scope, satisfy simultaneously.

5. a kind of pulverizing process fineness of pulverized coal index flexible measurement method according to claim 1 based on least square-support vector machine, the hardware platform core that it is characterized in that its dependence is made of the computer system that has been equipped with measurement instrument simultaneously and has carried out computed in software low speed coal mill, mill separator, cloth bag gatherer and relevant device; The input end of coal pulverizer and feeder, hot blast and cold air duct join, coal pulverizer outlet, mill separator and cloth bag gatherer link to each other successively, thermopair and thermal resistance thermometric original paper are housed respectively on hot blast and coal pulverizer outlet conduit, difference setting pressure meter on the import and export pipeline of coal pulverizer and cloth bag gatherer, feeder speed governing frequency converter output frequency feedback signal, system configuration control computer or data acquisition unit are gathered the measurement instrument signal; Survey Software operates on the supervisory control comuter of computer control system or independently on the computing machine.

Images