CN105278520A - Complex industrial process running state evaluation method and application based on T-KPRM - Google Patents

Abstract

The invention discloses a complex industrial process running state evaluation method and application based on T-KPRM. Combined with advantages of PRM and T-KPLS, the method comprises the steps of: further decomposing a high-dimension principal component subspace and a residual error subspace of KPLS, separating the part relevant to output from the part irrelevant to output, separating the part having a relatively large residual error from the part of final noise, and accurately extracting variable information relevant to the output. By adopting the method, the field industrial process running state can be conveniently grasped. According to the invention, an off-line evaluation module of the operation state is established, a sliding window technology is introduced, and the similarity between an online data window and corresponding evaluation grades is utilized for carrying out complex industrial process operation state online evaluation; in addition, according to an Euclidean distance between the sliding data window and an optimal evaluation grade, contribution rates of corresponding variables are calculated, and non-optimum factors of the operation state are identified, so that filed operators can adjust and improve the production strategy in time, and the production efficiency is improved.

Description

Based on T-KPRM complex industrial process evaluation of running status method and application

Technical field

The present invention relates to a kind of industrial process evaluation of running status method based on T-KPRM and application thereof, belong to industrial processes evaluation of running status technical field.

Background technology

Good industrial process running status is effective guarantee of enterprise product quality and economic benefit.But in complex industrial process, running status can be subject to the impact of various uncertain factor often, and causes it to depart from optimum operating point.Meanwhile, from the data of industry spot, in gatherer process, also can be subject to various interference (as sensor fault etc.), and cause there is outlier in sample data.Therefore, the industry park plan state on-line evaluation strategy proposing a kind of robust is necessary for grasping the ruuning situation of complex industrial in real time.And utilize dense medium to carry out the sorting of raw coal, a complicated industry park plan process just.Therebetween, need constantly to carry out measuring and adjustation to coal preparation technique parameter, to ensure the stable of product quality and quantity, also will ensure that the safety of production run is carried out simultaneously.

Running status on-line evaluation refers under commercial production remains on the normal condition run, and makes further differentiation to the good and bad situation of production run.Operating personnel according to evaluation result, can carry out production adjustment, to guarantee that production run process is in optimum state all the time.

No matter dense medium coal cleaning process, as domestic application coal preparation technique more widely, is dual medium cyclone dressing or gravity dense medium separation technique, and carrying out observing and controlling to Media density is the key ensureing product quality.To obtain good separating effect, just must ensure that Media density can be stabilized in a comparatively ideal scope.By dense medium coal separation density observing and controlling process operation state on-line evaluation, site operation personnel can be allowed to understand the concrete condition of Media density observing and controlling process operation state in time.Then according to evaluation result, corresponding adjustment is made to coal preparation technique, improve the efficiency of separation of coal separation process, increase the output capacity of cleaned coal, thus improve the economic benefit of enterprise.In addition, the raising of the efficiency of separation also contributes to development clean energy resource and recycling economy.

Dense medium cyclone is utilized to carry out the technological process of raw coal separation as shown in Figure 1, wherein, raw coal sends into mixing tank together with circulatory mediator, sorting is carried out again through two product heavy-medium cyclones, its underflow and overflow carry out de-mediated, dehydration and relevant treatment respectively, finally obtain cleaned coal, middle coal and spoil product, medium is then undertaken reclaiming, concentrating by medium system, and finally recycles.Whole assorting room is divided into coal stream flow process and medium flow process etc., and the sensor related in technique mainly contains densitometer, magnetic-substance content meter, liquid level gauge etc.Utilize the density in the involutory Jie's bucket of densitometer to monitor, utilize liquid level gauge to monitor the liquid level of closing suspending liquid in Jie's bucket; Utilize magnetic-substance content meter to monitor the coal slime in assorting room, this external system is equipped with all kinds of solenoid valve and is used for controlling the flow such as medium, clear water.Major control object in dense medium assorting room is close the heavy-media suspension density in Jie's bucket, and the control of density relates to Density feedback and ash content feeds back, the density of suspending liquid that Density feedback is used in involutory Jie's bucket carries out closed loop adjustment, ash content feedback is used for monitoring sorted product quality, adjust according to ash content result dense medium density of suspending liquid set-point, to reach the object of stabilized product quality.Coal preparation plant fast ash chemical examination detect be former pit ash degree, grey calibration is the basic index of Coal Quality.Ash calibration is less, and show that the coal impurity that washing goes out is fewer, Coal Quality is better.

At present, existing relevant scholar has done some researchs to production run evaluation of running status.To the research of dense-medium separation density observing and controlling process aspect, as far away in South Korea etc., Ding Yanjuan etc., proposed method only rests on density TT&C system on the theoretical analysis of whole dense-medium separation process importance, does not propose concrete density monitoring strategies.Liu Yan etc., propose a kind of evaluation of running status strategy based on T-KPLS algorithm.But, the data of industry spot are nonlinear often, but also include the outlier of some, the evaluation model using existing method (as T-KPLS, Fisher discriminatory analysis) to set up easily is subject to the impact of outlier, lose the due generalization ability of the method, have impact on the accuracy of evaluation result.

Summary of the invention

For above-mentioned prior art Problems existing, the invention provides a kind of complex industrial process evaluation of running status method based on T-KPRM, not only can solve the nonlinear problem of data rapidly, and can the information of each variable in leaching process data exactly, the impact of outlier on evaluation model precision can also be overcome; The application of a kind of complex industrial process evaluation of running status method (T-KPRM) based on T-KPRM provided refers to the complex industrial process evaluation of running status method based on core inclined robust latent variable technology, namely utilizes core inclined robust latent variable technology the method to can be used for setting up the evaluated off-line model of running status; Then, introduce sliding window technique, calculate the similarity between online data window and corresponding opinion rating, utilize similarity to carry out the on-line evaluation of complex industrial process running status.Simultaneously, Euclidean distance between slip data window and optimum opinion rating can be utilized, calculate the contribution rate of relevant variable, the non-optimal factor of running status is identified, there is provided decision-making foundation for site operation personnel adjusts production strategy in real time, thus improve the production efficiency of enterprise.

To achieve these goals, a kind of complex industrial process evaluation of running status method based on T-KPRM that the present invention adopts, supposes that process data matrix is X ∈ R ^{n × J}, N is sample number, and J is process variable number; Exporting data matrix is Y ∈ R ⁿ, comprise output procedure variable; Concrete steps then based on the complex industrial process evaluation of running status method of core inclined robust latent variable technology are as follows:

Step one, zero-mean and unit variance process are carried out to each row of input data matrix X; In like manner, also standardization is carried out to output data matrix Y;

Step 2, by input data matrix X through Nonlinear Mapping Φ: x _i∈ R ⁿ→ Φ (x _i) ∈ F projects to high-dimensional feature space F, and calculate nuclear matrix K:K=Φ in F space ^tΦ;

Step 3, standardization is carried out to nuclear matrix K;

Step 4, PRM algorithm is run to input nucleus matrix K and output matrix Y, specific as follows:

A1, to establish be the lever weights of i-th sample data, can be represented by following formula:

$w_i^x=f(\frac{||t_i-{\mathrm{med}}_{L1}\left(T\right)||}{{\mathrm{med}}_i||t_i-{\mathrm{med}}_{L1}\left(T\right)||},c),i=1,2,...n---\left(1\right),$

And $f(z,c)=\frac{1}{{(1+|z/c\left|\right)}^2}---\left(2\right),$

Wherein, || || represent Euclidean distance, med represents median, med _l1represent L1 median, t _ibe the PLS score of i-th sample data, c is constant;

If the residual error weights of i-th sample data can be defined by following formula:

Wherein, r _irepresent the residual error between i-th sample data predicted value and actual value, represent the robust size estimation of residual error, can be calculated by following formula: $\tilde{r}={\mathrm{med}}_i|r_i-{\mathrm{med}}_j\left(r_j\right)|,i,j=1,2,...n---\left(4\right);$

The then comprehensive weight w of i-th sample data _i, can be determined by following formula: utilize formula (1), (3), (5) initialization weights W _i;

B1, respectively input nucleus matrix K and output matrix Y to be weighted, after weighting, to obtain input nucleus matrix K _wwith output matrix Y _w, then PLS1 regression model is set up to the inputoutput data after weighting, revises its score vector simultaneously;

C1, calculate the residual error r of each data sample _i, utilize formula (1), (3), (5) to upgrade sample weights W _i;

If the relative difference of the double q of calculating in d1 front and back is less than threshold value (as 10 ^-2), then enter step 5, otherwise return b1;

Step 5, now input data matrix become K _w, export data matrix and become Y _w, from output matrix Y _wthe middle u extracting convergence _i, make i=1, K _wi=K _w, Y _wi=Y _w;

A2, make Y _wiin any one row equal u _i;

B2, calculating K _wscore vector, t _i=K _wiu _i, t _i← t _i/ || t _i||;

c2、 $q_i=Y_{Wi}^T{t}_i;$

D2, calculating Y _wiscore vector u _i=Y _wiq _i, u _i← u _i/ || u _i||;

E2, judge u _iwhether restrain, if convergence, proceed to step 6, otherwise return a2;

Step 6, calculating K _wiload matrix:

Step 7, extract whole pivot, calculate input data matrix K _wscore matrix T, input data matrix K _wload matrix P, export data matrix Y _wscore matrix U, export data matrix Y _wload matrix Q, specific as follows:

A3, order $K_{Wi+1}=K_{Wi}-t_i{p}_i^T;Y_{Wi+1}=Y_{Wi}-u_i{q}_i^T;$

B3, make i=i+1, repeat step 5, six until extract A pivot, pivot number A can be determined by cross-validation method;

c3、T＝[t ₁,...,t _A]，P＝[p ₁,...,p _A]，U＝[u ₁,...,u _A]，Q＝[q ₁,...,q _A]；

Step 8, K _w=TP ^t+ E, Y _w=UQ ^t+ F;

Step 9, to pivot TP ^trun PCA algorithm:

Step 10, PCA algorithm is run to residual error E:

If exporting data matrix Y is single output variable, then the expression formula of T-KPRM model is as follows:

$\left\{\begin{array}{c}{K}_W=t_y{p}_y^T+T_o{P}_o^T+T_r{P}_r^T+E_r\\ y_W=t_y+F\end{array}\right.---\left(6\right)$

Above-mentioned score vector t _ythat represent is in T and output y _wdirectly related part, namely sets up the most critical variable drawn required for evaluation model, can be used for the foundation of running status evaluated off-line model; T _orepresent in T with output y _worthogonal part; ; T _rit is the part having larger variance in residual error E; E _rfinal residual error, namely noise;

When introducing new sample k _new, its score matrix will be drawn by following formula:

$t_y=y_W^{T}T{\left(U^T{K}_{W}T\right)}^{-1}{U}^T{k}_{new}=g_y{k}_{new}---\left(7\right),$

Wherein, k _newnew samples x _newkernel function, can be calculated by following formula:

k _new＝Φ(X)Φ(x _new)＝[k(x ₁,x _new),...k(x _n,x _new)] ^T(8)，

To k _newcarry out equalization can obtain: ${\stackrel{\‾}{k}}_{new}=k_{new}-K{1}_t-1_n{k}_{new}+1_{n}K{1}_t---\left(9\right),$

Wherein, 1 _t=1/n [11...1] ^t∈ R ⁿ.Above-mentioned score vector t _ycan be used for the foundation of running status evaluated off-line model.

A kind of application of the complex industrial process evaluation of running status method based on T-KPRM, specifically refer to the identification utilizing core inclined robust latent variable technology to set up the evaluated off-line model of running status, the on-line evaluation of complex industrial process running status and non-optimal factor, specifically first utilize core inclined robust latent variable technology to set up evaluated off-line model, the evaluated off-line model score vector of corresponding running status grade is wherein, c is the number of running status grade; Then, introduce sliding window technique, calculate the similarity between online data window and corresponding opinion rating, utilize similarity to carry out the on-line evaluation of complex industrial process running status; Euclidean distance between recycling slip data window and optimum opinion rating, calculates the contribution rate of relevant variable, identifies the non-optimal factor of running status.

Carrying out in evaluation of running status process, consider that single sample data is not enough to characterize the running status of production run, but also be easily subject to the interference of various factors, therefore, introducing a width is that the data window of H is as basic analytic unit; Then, the similarity between k time data window and corresponding opinion rating is calculated for carrying out the evaluation of process operation state; Meanwhile, a similarity threshold ε (0.5 < ε < 1) is set, for distinguishing the transition period of running status grade and the corresponding running status grade determined; Suppose the maximal value of similarity larger than threshold epsilon, then can deterministic process state grade be p; When running status is in transition period, time namely from the stage that a upper running status changes to next running status, the value of its similarity presents a rule slowly changed; Suppose that similarity is less than threshold epsilon and its size meets the rule increased progressively in turn, then can think that process operation state is in a transition period; Be enough to two hypothesis if discontented, just can think that running status is disturbed by other factors, result of its evaluation and being consistent before.

The flow process of on-line operation state evaluation is as follows:

The first step: the line window data matrix in structure k moment, X _{on, k}=[x _{on, k-H+1}..., x _{on, k}] ^t, H is the width of data window;

Second step: standardization X _{on, k}, the evaluated off-line model established before utilization, the line window data matrix after asking for standardization, ${\stackrel{\&OverBar;}{X}}_{on,k}^c,c=1,2,3,4;$

3rd step: calculate score vector

Wherein, by $k_{on,k}^c={\&lsqb;k(x_1^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c),k(x_2^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c),...,k(x_{n_c}^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c)\&rsqb;}^T$ Average centralization gained;

4th step: calculate the Euclidean distance between online data window and corresponding opinion rating c, wherein be equalization vector; Via T-KPRM method gained, then

5th step: utilize value, ask for the similarity between data window and corresponding opinion rating c:

Suppose $d_{on,k}^c\&NotEqual;0,$ Then ${\mathrm{\&gamma;}}_{on,k}^c=\frac{1/d_{on,k}^c}{\underset{p=1}{\overset{4}{\&Sigma;}}1/d_{on,k}^p}---\left(10\right),$ Wherein $\underset{c=1}{\overset{4}{\&Sigma;}}{\mathrm{\&gamma;}}_{on,k}^c=1,0\&le;{\mathrm{\&gamma;}}_{on,k}^c\&le;1;$

6th step: judge running status grade according to similarity, the interpretational criteria of on-line operation state is as follows:

If criterion one then judge that this online data window belongs to some running status grade p determined;

If criterion two does not meet criterion one, and meets formula $p=\mathrm{argmax}\left\{{\mathrm{\&gamma;}}_{on,k}^p\right|{\mathrm{\&gamma;}}_{on,k-l+1}^p<...<{\mathrm{\&gamma;}}_{on,k}^p,1\&le;p\&le;4\},$ Then judge that this data window is in the transition period of running status, be namely in the stage transformed to next running status from a upper running status, wherein, l is a positive integer, and size is determined by knowhow;

If criterion three is discontented be enough to two criterions, then prove that this data window is subject to the interference of other factors, the judged result of its running status be consistent before;

By the judgement of above three criterions, just can effectively determine to belong to some running status grades or the transition period be between state grade at the data window in k moment.

Non-optimal running status refers to all running statuses outside removing optimal operational condition, and this is comprising kilter, the transition period between general state, poor state and steady state (SS).Generally speaking, the supvr of enterprise wishes that production run runs on best state.But, when commercial production is inevitably in the non-optimal stage, goes to review those and cause running with regard to needing to go for a kind of suitable method the reason being in non-optimal conditions, thus provide foundation effectively for production adjustment.

In the research of the fault diagnosis based on data, the utilization of contribution plot method is commonplace.Particularly be directed to linear model, the method intuitively can find out the correlated variables that those causing trouble occur effectively.But traditional contribution plot method is not also suitable for the model set up by kernel function.Proposed here a kind of recognition methods of the non-optimal factor based on kernel function: first by calculating the partial derivative of Euclidean distance between online data window and optimal operational condition grade, obtain the contribution rate of sample data variable, the variable of larger contribution rate is wherein had to cause producing by being considered to the factor being in non-optimal running status, then to these factors Discern and judge in addition.

Euclidean distance between online data window and optimal operational condition grade is expressed as again

$d_{on,k}^4=t_{on,k}^{4T}{t}_{on,k}^4={\stackrel{\&OverBar;}{k}}_{on,k}^{4T}{g}_y^{4T}{g}_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4=tr\left(g_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4\text{T}}{g}_y^{4T}\right)---\left(11\right),$

In formula, represent the off-line model matrix of optimal operational condition, represent the core Mean Matrix of the online data window of corresponding optimal operational condition;

? in formula, the contribution rate of a jth variable be can be calculated by following formula:

$\begin{array}{c}{\mathrm{Contr}}_j=\left|\frac{\&part;d_{on,k}^4}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\right|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4\\ =\left|\frac{\&part;}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}tr\left(g_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4T}{g}_y^{4T}\right)\right|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4\\ =|tr\&lsqb;g_y^4\left(\frac{\&part;{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4T}}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\right)g_y^{4T}\&rsqb;|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4,j=1,2,...,m,\end{array}---\left(12\right),$

In formula, it is sample data window a jth variable, tr () is expressed as matrix trace;

In order to try to achieve above formula Contr _jexpression formula accurately, must draw partial derivative, therefore, need the process carrying out some necessity:

A, nuclear matrix partial derivative be: $\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}=\frac{1}{\mathrm{\&sigma;}}(x_{i,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)---\left(13\right),$

B, partial derivative is asked to two nuclear matrix simultaneously, can obtain:

$\begin{array}{c}\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)+k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\frac{\&part;k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{1}{\mathrm{\&sigma;}}(x_{i,j}+x_{q,j}-2{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\end{array}---\left(14\right),$

C, to calculate in the partial derivative of element that arranges of the i-th row q:

$\begin{array}{c}\begin{array}{c}\frac{\&part;{({\stackrel{\&OverBar;}{k}}_{on,k}^4,{\stackrel{\&OverBar;}{k}}_{on,k}^{4T})}_{pq}}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}=\frac{\&part;(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +(A-a_q)\frac{\&part;(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}+(A-a_p)\frac{\&part;(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ -\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ -\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +\frac{(a_p+q_q-2A)}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +\frac{1}{n_4^2}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\underset{i^{\&prime;}=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_{i^{\&prime;}},{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{1}{\mathrm{\&sigma;}}(x_{p,j}+x_{q,j}-{2\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(A-a_q)}{\mathrm{\&sigma;}}(x_{p,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(A-a_p)}{\mathrm{\&sigma;}}(x_{q,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ -\frac{1}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{p.j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ -\frac{1}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{q.j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(a_p+a_q-2A)}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}-{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{1}{{\mathrm{\&sigma;n}}_4^2}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\underset{i^{\&prime;}=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{i^{\&prime;},j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_{i^{\&prime;}},{\stackrel{\&OverBar;}{x}}_{on,k}^4)\end{array}\end{array}---\left(15\right),$

In formula, $a_p=\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\&Sigma;}}k(x_p,x_i),A=\frac{1}{n_4^2}\underset{i=1}{\overset{n_4}{\&Sigma;}}\underset{i^{\&prime;}=1}{\overset{n_4}{\&Sigma;}}k(x_i,x_{i^{\&prime;}}),$ N ₄it is the sample number of the modeling data of optimal operational condition.

Compared with prior art, the complex industrial process evaluation of running status method (T-KPRM) based on core inclined robust latent variable technology that the present invention proposes combines the advantage of both PRM and T-KPLS, can grasp the running status of situ industrial process real-time and accurately, T-KPPM can decompose higher-dimension principal component subspace and the residual error subspace of KPLS further.In principal component subspace, the part irrelevant with output with exporting relevant part is separated.Meanwhile, in residual error subspace, separate there being the part of the part of larger residual error and final noise.Therefore, T-KPRM can extract exactly and export relevant variable information.The method can be used for the evaluated off-line model setting up running status.Meanwhile, introduce sliding window technique, calculate the similarity between online data window and corresponding opinion rating, similarity can be utilized to carry out the on-line evaluation of complex industrial process running status.Afterwards, utilize Euclidean distance between slip data window and optimum opinion rating, calculate the contribution rate of relevant variable, the non-optimal factor of running status is identified.Contrasted by simulation analysis, the evaluated off-line model of the running status utilizing the complex industrial process evaluation of running status method (T-KPRM) based on core inclined robust latent variable technology to set up, dense medium coal separation density observing and controlling process carries out running status on-line evaluation and the identification of non-optimal factor is of practical significance very much.By means of evaluating and the result of factor identification, operating personnel just can adjust in time and improve production strategy, improve production efficiency.In addition, this running status on-line evaluation method is all applicable to dual medium cyclone dressing and gravity dense medium separation technique.

Accompanying drawing explanation

Fig. 1 is the process chart utilizing dense medium cyclone to carry out raw coal separation;

Fig. 2 is the on-line operation state evaluation and the non-optimal factor identification process figure that the present invention is based on T-KPRM;

Fig. 3 is similarity and density observing and controlling running status on-line evaluation result schematic diagram;

Fig. 4 is existing method (T-KPLS) the density observing and controlling evaluation of running status result not having very noisy and outlier;

Fig. 5 is not for having institute's extracting method (T-KPRM) density observing and controlling evaluation of running status result of very noisy and outlier;

Fig. 6 is existing method (T-KPLS) the density observing and controlling process operation state evaluation result having very noisy and outlier;

Fig. 7 is for there being institute's extracting method (T-KPRM) density observing and controlling evaluation of running status result of very noisy and outlier;

Fig. 8 is the non-optimal factor identification of state grade " poor ";

Fig. 9 be " poor " arrive " in " between the non-optimal factor identification of transition period;

Figure 10 be state grade " in " the identification of non-optimal factor;

Figure 11 be " in " arrive the non-optimal factor identification of transition period between " good ";

Figure 12 is the non-optimal factor identification of state grade " good ";

Figure 13 is the non-optimal factor identification that " good " arrives transition period between " excellent ".

Embodiment

Below in conjunction with accompanying drawing and simulation analysis, the invention will be further described.

Based on a complex industrial process evaluation of running status method of T-KPRM, suppose that process data matrix is X ∈ R ^{n × J}, N is sample number, and J is process variable number; Exporting data matrix is Y ∈ R ⁿ, comprise output procedure variable; Concrete steps then based on the complex industrial process evaluation of running status method of core inclined robust latent variable technology are as follows:

Step one, zero-mean and unit variance process are carried out to each row of input data matrix X; In like manner, also standardization is carried out to output data matrix Y;

Step 2, by input data matrix X through Nonlinear Mapping Φ: x _i∈ R ⁿ→ Φ (x _i) ∈ F projects to high-dimensional feature space F, and calculate nuclear matrix K:K=Φ in F space ^tΦ;

Step 3, standardization is carried out to nuclear matrix K;

Step 4, PRM algorithm is run to input nucleus matrix K and output matrix Y, specific as follows:

A1, to establish be the lever weights of i-th sample data, can be represented by following formula:

$w_i^x=f(\frac{||t_i-{\mathrm{med}}_{L1}\left(T\right)||}{{\mathrm{med}}_i||t_i-{\mathrm{med}}_{L1}\left(T\right)||},c),i=1,2,...n---\left(1\right),$

And $f(z,c)=\frac{1}{{(1+|z/c\left|\right)}^2}---\left(2\right),$

Wherein, || || represent Euclidean distance, med represents median, med _l1represent L1 median, t _ibe the PLS score of i-th sample data, c is constant;

If the residual error weights of i-th sample data can be defined by following formula:

Wherein, r _irepresent the residual error between i-th sample data predicted value and actual value, represent the robust size estimation of residual error, can be calculated by following formula: $\tilde{r}={\mathrm{med}}_i|r_i-{\mathrm{med}}_j\left(r_j\right)|,i,j=1,2,...n---\left(4\right);$

The then comprehensive weight w of i-th sample data _i, can be determined by following formula: utilize formula (1), (3), (5) initialization weights W _i;

B1, respectively input nucleus matrix K and output matrix Y to be weighted, after weighting, to obtain input nucleus matrix K _wwith output matrix Y _w, then PLS1 regression model is set up to the inputoutput data after weighting, revises its score vector simultaneously;

C1, calculate the residual error r of each data sample _i, utilize formula (1), (3), (5) to upgrade sample weights W _i;

If the relative difference of the double q of calculating in d1 front and back is less than threshold value (as 10 ^-2), then enter step 5, otherwise return b1;

Step 5, now input data matrix become K _w, export data matrix and become Y _w, from output matrix Y _wthe middle u extracting convergence _i, make i=1, K _wi=K _w, Y _wi=Y _w;

A2, make Y _wiin any one row equal u _i;

B2, calculating K _wscore vector, t _i=K _wiu _i, t _i← t _i/ || t _i||;

c2、 $q_i=Y_{Wi}^T{t}_i;$

D2, calculating Y _wiscore vector u _i=Y _wiq _i, u _i← u _i/ || u _i||;

E2, judge u _iwhether restrain, if convergence, proceed to step 6, otherwise return a2;

Step 6, calculating K _wiload matrix:

Step 7, extract whole pivot, calculate input data matrix K _wscore matrix T, input data matrix K _wload matrix P, export data matrix Y _wscore matrix U, export data matrix Y _wload matrix Q, specific as follows:

A3, order $K_{Wi+1}=K_{Wi}-t_i{p}_i^T;Y_{Wi+1}=Y_{Wi}-u_i{q}_i^T;$

B3, make i=i+1, repeat step 5, six until extract A pivot, pivot number A can be determined by cross-validation method;

c3、T＝[t ₁,...,t _A]，P＝[p ₁,...,p _A]，U＝[u ₁,...,u _A]，Q＝[q ₁,...,q _A]；

Step 8, K _w=TP ^t+ E, Y _w=UQ ^t+ F;

Step 9, to pivot TP ^trun PCA algorithm:

Step 10, PCA algorithm is run to residual error E:

If exporting data matrix Y is single output variable, then the expression formula of T-KPRM model is as follows:

$\left\{\begin{array}{c}{K}_W=t_y{p}_y^T+T_o{P}_o^T+T_r{P}_r^T+E_r\\ y_W=t_y+F\end{array}\right.---\left(6\right)$

Above-mentioned score vector t _ythat represent is in T and output y _wdirectly related part, namely sets up the most critical variable drawn required for evaluation model, can be used for the foundation of running status evaluated off-line model; T _orepresent in T with output y _worthogonal part; ; T _rit is the part having larger variance in residual error E; E _rfinal residual error, namely noise;

When introducing new sample k _new, its score matrix will be drawn by following formula:

$t_y=y_W^{T}T{\left(U^T{K}_{W}T\right)}^{-1}{U}^T{k}_{new}=g_y{k}_{new}---\left(7\right),$

Wherein, k _newnew samples x _newkernel function, can be calculated by following formula:

k _new＝Φ(X)Φ(x _new)＝[k(x ₁,x _new),...k(x _n,x _new)] ^T(8)，

To k _newcarry out equalization can obtain: ${\stackrel{\&OverBar;}{k}}_{new}=k_{new}-K{1}_t-1_n{k}_{new}+1_{n}K{1}_t---\left(9\right),$

Wherein, 1 _t=1/n [11...1] ^t∈ R ⁿ.Above-mentioned score vector t _ycan be used for the foundation of running status evaluated off-line model.

T-KPRM, in conjunction with the advantage of both PRM and T-KPLS, can grasp the running status of situ industrial process real-time and accurately.T-KPPM can decompose higher-dimension principal component subspace and the residual error subspace of KPLS further.In principal component subspace, the part irrelevant with output with exporting relevant part is separated.Meanwhile, in residual error subspace, separate there being the part of the part of larger residual error and final noise.Therefore, T-KPRM can extract exactly and export relevant variable information.

A kind of application of the complex industrial process evaluation of running status method based on T-KPRM, specifically refer to the identification utilizing core inclined robust latent variable technology to set up the evaluated off-line model of running status, the on-line evaluation of complex industrial process running status and non-optimal factor, specifically first utilize core inclined robust latent variable technology to set up evaluated off-line model, the evaluated off-line model score vector of corresponding running status grade is wherein, c is the number of running status grade; Then, introduce sliding window technique, calculate the similarity between online data window and corresponding opinion rating, utilize similarity to carry out the on-line evaluation of complex industrial process running status; Euclidean distance between recycling slip data window and optimum opinion rating, calculates the contribution rate of relevant variable, identifies the non-optimal factor of running status.

Carrying out in evaluation of running status process, consider that single sample data is not enough to characterize the running status of production run, but also be easily subject to the interference of various factors, therefore, introducing a width is that the data window of H is as basic analytic unit; Then, the similarity between k time data window and corresponding opinion rating is calculated for carrying out the evaluation of process operation state; Meanwhile, a similarity threshold ε (0.5 < ε < 1) is set, for distinguishing the transition period of running status grade and the corresponding running status grade determined; Suppose the maximal value of similarity larger than threshold epsilon, then can deterministic process state grade be p; When running status is in transition period, time namely from the stage that a upper running status changes to next running status, the value of its similarity presents a rule slowly changed; Suppose that similarity is less than threshold epsilon and its size meets the rule increased progressively in turn, then can think that process operation state is in a transition period; Be enough to two hypothesis if discontented, just can think that running status is disturbed by other factors, result of its evaluation and being consistent before.

The flow process of on-line operation state evaluation is as follows:

The first step: the line window data matrix in structure k moment, X _{on, k}=[x _{on, k-H+1}..., x _{on, k}] ^t, H is the width of data window;

Second step: standardization X _{on, k}, the evaluated off-line model established before utilization, the line window data matrix after asking for standardization, ${\stackrel{\&OverBar;}{X}}_{on,k}^c,c=1,2,3,4;$

3rd step: calculate score vector

Wherein, by $k_{on,k}^c={\&lsqb;k(x_1^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c),k(x_2^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c),...,k(x_{n_c}^c,{\stackrel{\&OverBar;}{x}}_{on,k}^c)\&rsqb;}^T$ Average centralization gained;

4th step: calculate the Euclidean distance between online data window and corresponding opinion rating c, wherein be equalization vector; Via T-KPRM method gained, then

5th step: utilize value, ask for the similarity between data window and corresponding opinion rating c:

Suppose $d_{on,k}^c\&NotEqual;0,$ Then ${\mathrm{\&gamma;}}_{on,k}^c=\frac{1/d_{on,k}^c}{\underset{p=1}{\overset{4}{\&Sigma;}}1/d_{on,k}^p}\text{---}\left(10\right),$ Wherein $\underset{c=1}{\overset{4}{\&Sigma;}}{\mathrm{\&gamma;}}_{on,k}^c=1,0\&le;{\mathrm{\&gamma;}}_{on,k}^c\&le;1;$

6th step: judge running status grade according to similarity, the interpretational criteria of on-line operation state is as follows:

If criterion one then judge that this online data window belongs to some running status grade p determined;

If criterion two does not meet criterion one, and meets formula $p=\mathrm{argmax}\left\{{\mathrm{\&gamma;}}_{on,k}^p\right|{\mathrm{\&gamma;}}_{on,k-l\text{+1}}^p<...<{\mathrm{\&gamma;}}_{on,k}^p,1\&le;p\&le;4\},$ Then judge that this data window is in the transition period of running status, be namely in the stage transformed to next running status from a upper running status, wherein, l is a positive integer, and size is determined by knowhow;

If criterion three is discontented be enough to two criterions, then prove that this data window receives the interference of other factors, the judged result of its running status be consistent before;

By the judgement of above three criterions, just can effectively determine to belong to some running status grades or the transition period be between state grade at the window data in k moment.

Non-optimal running status refers to all running statuses outside removing optimal operational condition, and this is comprising kilter, the transition stage of general state, poor state and state.Generally speaking, the supvr of enterprise wishes that production run runs on best state.But, when commercial production is inevitably in the non-optimal stage, goes to review those and cause running with regard to needing to go for a kind of suitable method the reason being in non-optimal conditions, thus provide foundation effectively for production adjustment.

In the research of the fault diagnosis based on data, the utilization of contribution plot method is commonplace.Particularly be directed to linear model, the method intuitively can find out the correlated variables that those causing trouble occur effectively.But traditional contribution plot method is not also suitable for the model set up by kernel function.Proposed here a kind of recognition methods of the non-optimal factor based on kernel function: first by calculating the partial derivative of Euclidean distance between online data window and optimal operational condition grade, obtain the contribution rate of sample data variable, the variable of larger contribution rate is wherein had to cause producing by being considered to the factor being in non-optimal running status, then to these factors Discern and judge in addition.

Euclidean distance between online data window and optimal operational condition grade is expressed as again

$d_{on,k}^4=t_{on,k}^{4T}{t}_{on,k}^4={\stackrel{\&OverBar;}{k}}_{on,k}^{4T}{g}_y^{4T}{g}_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4=tr\left(g_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4T}{g}_y^{4T}\right)---\left(11\right),$

In formula, represent the off-line model matrix of optimal operational condition, represent the core Mean Matrix of the online data window of corresponding optimal operational condition;

? in formula, the contribution rate of a jth variable be can be calculated by following formula:

$\begin{array}{c}{\mathrm{Contr}}_j=\left|\frac{\&part;d_{on,k}^4}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\right|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4\\ =\left|\frac{\&part;}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}tr\left(g_y^4{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4T}{g}_y^{4T}\right)\right|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4\\ =|tr\&lsqb;g_y^4\left(\frac{\&part;{\stackrel{\&OverBar;}{k}}_{on,k}^4{\stackrel{\&OverBar;}{k}}_{on,k}^{4T}}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\right)g_y^{4T}\&rsqb;|\&CenterDot;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4,j=1,2,...,m,\end{array}---\left(12\right),$

In formula, it is sample data window a jth variable, tr () is expressed as matrix trace;

In order to try to achieve above formula Contr _jexpression formula accurately, must draw partial derivative, therefore, need the process carrying out some necessity:

A, nuclear matrix partial derivative be: $\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}=\frac{1}{\mathrm{\&sigma;}}(x_{i,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)---\left(13\right),$

B, partial derivative is asked to two nuclear matrix simultaneously, can obtain:

$\begin{array}{c}\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{\&part;k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)+k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\frac{\&part;k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{1}{\mathrm{\&sigma;}}(x_{i,j}+x_{q,j}-2{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\end{array}---\left(14\right),$

C, to calculate in the partial derivative of element that arranges of the i-th row q:

$\begin{array}{c}\begin{array}{c}\frac{\&part;{({\stackrel{\&OverBar;}{k}}_{on,k}^4,{\stackrel{\&OverBar;}{k}}_{on,k}^{4T})}_{pq}}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}=\frac{\&part;(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +(A-a_q)\frac{\&part;(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}+(A-a_p)\frac{\&part;(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ -\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ -\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +\frac{(a_p+q_q-2A)}{n_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ +\frac{1}{n_4^2}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\underset{i^{\&prime;}=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\frac{\&part;(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_{i^{\&prime;}},{\stackrel{\&OverBar;}{x}}_{on,k}^4)}{\&part;{\stackrel{\&OverBar;}{x}}_{on,k,j}^4}\\ =\frac{1}{\mathrm{\&sigma;}}(x_{p,j}+x_{q,j}-{2\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(A-a_q)}{\mathrm{\&sigma;}}(x_{p,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(A-a_p)}{\mathrm{\&sigma;}}(x_{q,j}-{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ -\frac{1}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{p.j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_p,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ -\frac{1}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{q.j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_q,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{(a_p+a_q-2A)}{{\mathrm{\&sigma;n}}_4}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}-{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)\\ +\frac{1}{{\mathrm{\&sigma;n}}_4^2}\underset{i=1}{\overset{n_4}{\mathrm{\&Sigma;}}}\underset{i^{\&prime;}=1}{\overset{n_4}{\mathrm{\&Sigma;}}}(x_{i,j}+x_{i^{\&prime;},j}-2{\stackrel{\&OverBar;}{x}}_{on,k,j}^4)k(x_i,{\stackrel{\&OverBar;}{x}}_{on,k}^4)k(x_{i^{\&prime;}},{\stackrel{\&OverBar;}{x}}_{on,k}^4)\end{array}\end{array}---\left(15\right),$

In formula, $a_p=\frac{1}{n_4}\underset{i=1}{\overset{n_4}{\&Sigma;}}k(x_p,x_i),A=\frac{1}{n_4^2}\underset{i=1}{\overset{n_4}{\&Sigma;}}\underset{i^{\&prime;}=1}{\overset{n_4}{\&Sigma;}}k(x_i,x_{i^{\&prime;}}),$ N ₄it is the sample number of the modeling data of optimal operational condition.

Simulation analysis:

In heavy medium coal preparation process, need constantly to carry out measuring and adjustation to coal preparation technique parameter, to ensure the stable of product quality and quantity, and ensure that the safety of production run is carried out, therefore the measuring and adjustation of the density of heavy-media suspension and rheological characteristics is the emphasis of dense medium coal cleaning process parameter investigating.In actual production, the running status of Dense Media Density observing and controlling process usually can be subject to the impact of the factors such as procedure parameter drift and environmental perturbation, and the running status of off-target.Dense medium is utilized to carry out the sorting of raw coal, just a complicated industry park plan process.In order to grasp the running status of production in time, this method can be used for the dense-medium separation density observing and controlling process operation state on-line evaluation of Kernel-based methods data.

By considering on-the-spot actual production process to the in-depth analysis of dense-medium separation density observing and controlling process, 5 process variable and 1 output variable on-line evaluation for dense-medium separation density observing and controlling process operation state can be chosen.These 5 process variable are respectively: main separation density, main separation liquid level, main separation pressure, Magnetic Materials concentration, coal slurry content; 1 output variable is: grey calibration.Each input/output variable is as shown in table 1:

Table 1 input/output variable table

1) foundation of density observing and controlling process off-line evaluation model

According to the field data of certain coal preparation plant's density observing and controlling process, 9900 groups of samples can be gathered, for modeling and the classification of assessment of running status.According to long-term practical production experience, the density of the heavy medium observing and controlling process is divided into gradual change running status between 4 kinds of state grades and different brackets according to grey calibration (fast grey data), as shown in table 2:

Table 2 process status criteria for classifying

Wherein, modeling data collection (X ₁, y ₁), (X ₂, y ₂), (X ₃, y ₃), (X ₄, y ₄) respectively corresponding excellent in differ from 4 class running status grades.Meanwhile, the evaluation model of each running status grade of T-KPRM method establishment is utilized c=1,2,3,4.

2) based on the density observing and controlling process on-line operation state evaluation of T-KPRM

Can in actual production process, random acquisition 640 groups of test datas in addition, for running status on-line evaluation.As shown in table 3, test data comprises the excellent middle 3 kinds of gradual change running state data differed between 4 state grades and different brackets.Arranging similarity threshold is ε=0.8, and moving window width is H=4, step-length l=1, and the continuous incremental number of similarity is 3.Production run follows time series, according to difference, in, good, excellent order transition.

Table 3 each stage test sample book group number

As shown in Figure 3, (a), (b), (c) in Fig. 3, (d) four figure represent respectively slip data window and difference, in, the situation of change of similarity between good, excellent one of four states grade.Between the 1st to the 103rd sampling instant, the similarity of the data of time slip-window and grade " poor " is greater than similarity threshold 0.85, can be judged by interpretational criteria one, and now production run operates in " poor " state.In like manner, the 127th to the 355th sampling instant, the 380th to the 509th sampling instant, between the 529th to the 640th sampling instant, the data of online data window respectively with state " in ", " good ", the similarity of " excellent " is all greater than similarity threshold 0.85, then can determine that now production run operates in respectively " in ", " good " and " excellent " state.(e) figure of Fig. 3 is the running status on-line evaluation result of whole observing and controlling process, that the ordinate 1,2,3,4 in figure represents respectively is poor, in, good, excellent one of four states grade.

Removing four determines outside the running status of grade, and remaining sampling instant region is then the transition period between adjacent two steady state (SS)s.Such as, from employing the moment 104 to 106 between, the similarity between line window data and each state grade is all less than threshold value, and with " in " similarity increase progressively gradually.Namely meet judge according to interpretational criteria two, meeting incremental number before and after similarity is 3, then show between sampling instant 104 to 126, the data of time slip-window and the similarity of " poor " grade are all less than similarity threshold, and the similarity between the moment of front and back presents the trend of successively decreasing successively.Production run run now be in " poor " arrive " in " transition period.In like manner, also can be judged as between sampling instant 356 to 379,510 to 528, production run is in respectively " in " arrive the transition period that the transition period of " good " and " good " arrive " excellent ".In (e) figure of Fig. 3, ordinate 1.5,2.5 and 3.5 represents the transition period between running status respectively.

3) evaluation model contrast when very noisy and outlier is had

Consider that industrial field data often can be subject to various interference (as sensor fault, process disturbance etc.) in gatherer process, and cause there is outlier in sample data.The existence of these outlier often causes modeling and analysis methods to lose its due generalization ability, thus affects the precision of model.Based on the complex industrial process evaluation of running status method (T-KPRM) of core inclined robust latent variable technology owing to combining the advantage of PRM and T-KPLS two kinds of algorithms, not only can solve the nonlinear problem of data rapidly, the impact of outlier on evaluation model precision can also be overcome.Although existing method (as T-KPLS, FDA etc.) also may be used for the foundation of evaluation of running status model, but when including outlier in modeling sample data, the precision of its evaluation model will be affected, and cannot carry out evaluation and the identification of running status exactly.

For dense-medium separation density observing and controlling process, random acquisition 603 groups of test datas in addition, for verifying the robustness that T-KPRM method is good.As shown in table 4, still comprise the excellent middle 3 kinds of gradual change running state data differed between 4 state grades and different brackets in these test datas.And four groups that modeling data collection still adopts one to save, be (X respectively ₁, y ₁), (X ₂, y ₂), (X ₃, y ₃), (X ₄, y ₄).Utilize the evaluation model of each running status grade of T-KPRM method establishment simultaneously c=1,2,3,4.Existing method (T-KPLS) is utilized also to set up the evaluation model of each running status grade c=1,2,3,4.

The test data that table 4 contrasts for method

Here select the mode adding disturbance in modeling data to simulate outlier and very noisy.Consider the collection situation of production data reality, " excellent " is compared with the modeling data of " good " grade " in " and the modeling data of " poor " grade to lack, therefore in the modeling data of " excellent " and " good " grade, the sample data of random choose 5% adds disturbance respectively.The mode that disturbance adds: by selected 5% sample (outlier), be divided into two parts data, and add the disturbance of 30% to the input of a part of data, the output of another part data of reciprocity quantity adds the disturbance of 30%.Meanwhile, " in " and " poor " grade modeling data in the sample data of random choose 10% respectively add disturbance.The mode that disturbance adds: by selected 10% sample (outlier), be divided into two parts data, and add the disturbance of 30% to the input of a part of data, the output of another part data of reciprocity quantity adds the disturbance of 30%.

When modeling data is normal data, when namely not adding disturbance, existing method (T-KPLS) and institute extracting method (T-KPRM) is utilized to carry out density observing and controlling process operation state evaluation, the evaluation result contrast of two kinds of methods, as shown in Figure 4, Figure 5.

When including outlier and very noisy in modeling data, namely according to statement before, add disturbance, existing method (T-KPLS) and institute extracting method (T-KPRM) is utilized to carry out density observing and controlling process operation state evaluation, the evaluation result contrast of two kinds of methods, as shown in Figure 6, Figure 7.

Below the evaluation recognition accuracy of two kinds of methods is contrasted.As shown in table 5, table 6, wherein similarity threshold span is: 0.6 to 0.9.

The existing method of table 5 (T-KPLS) evaluates recognition accuracy

Similarity threshold	Difference	In	Good	Excellent
					ε≥0.6	21％	13％	59％	12％
ε≥0.7	17％	9％	41％	6％
					ε≥0.8	10％	7％	19％	4％
ε≥0.9	4％	2％	11％	3％

Table 6 extracting method (T-KPRM) evaluates recognition accuracy

Similarity threshold	Difference	In	Good	Excellent
					ε≥0.6	99％	72％	99％	99％
ε≥0.7	99％	55％	98％	96％
					ε≥0.8	96％	30％	79％	87％
ε≥0.9	88％	15％	9％	62％

Analyze as can be seen from table, the evaluation recognition accuracy of T-KPRM method obviously will be better than the evaluation recognition accuracy of existing method (T-KPLS method).

Therefore, the comparing result of comprehensive evaluation model figure and recognition accuracy is visible, owing to including given outlier and very noisy in modeling data, existing method (T-KPLS) loses original generalization ability, the evaluation result figure of its four class states set up, all lose evaluation precision, and higher evaluation recognition accuracy cannot be kept, illustrate that the noise resisting ability of original method (T-KPLS) is weak.And the evaluation model that institute extracting method (T-KPRM) sets up, although there is some a little disturbances, but the emulation profile of evaluation result under substantially maintaining noise-free case, and maintain higher evaluation recognition accuracy, illustrate that the noise resisting ability of T-KPRM method is strong, method robustness is also good than existing method (T-KPLS).

4) for the non-optimal factor identification of density observing and controlling process

Non-optimal conditions refers to that industry park plan state is in other steady state (SS)s outside " excellent " state or transition state.When dense-medium separation density observing and controlling process operation is in some non-optimal conditions determined or the transition period being in former and later two steady state (SS)s, non-optimal factor now remains unchanged.That is, the change of non-optimal factor can cause different evaluation results.

To the non-optimal factor identification simulation result of each running status of dense-medium separation density observing and controlling process, as shown in Fig. 8,9,10,11,12,13, what its ordinate of Fig. 8 to Figure 13 represented is contribution rate, 1 in horizontal ordinate, 2,3,4,5 corresponding five variablees respectively: main separation density, main separation liquid level, main separation pressure, Magnetic Materials concentration and coal slurry content.The sampling instant that identifying is got is all random choose from the transition period between each state grade and state grade.Fig. 8 shows that the contribution rate of now main separation density and main separation pressure is higher, and therefore causing the 50th sampling instant to be in non-optimal conditions principal element is main separation density and main separation pressure.As can be seen from Figure 9 main separation pressure and main separation liquid level are the principal elements that the 115th sampling instant is in non-optimal conditions.From Figure 10,11,12,13, in like manner also can show that it is in the principal element of non-optimal conditions, does not repeat them here respectively.

Can be found out by above simulation result, utilize the running status evaluated off-line model that the complex industrial process evaluation of running status method (T-KPRM) based on core inclined robust latent variable technology is set up, dense medium coal separation density observing and controlling process carries out evaluation of running status and the identification of non-optimal factor is of practical significance very much.By means of evaluating and the result of factor identification, site operation personnel just can adjust in time and improve production strategy, enhances productivity.This evaluation of running status method is all applicable to dual medium cyclone dressing and gravity dense medium separation technique.

In sum, the T-KPRM method that the present invention proposes combines the advantage of both PRM and T-KPLS, effectively can not only solve the nonlinear problem of density observing and controlling process data, effectively can also overcome the impact of outlier on evaluation model precision.For complex industrial production run evaluation of running status problem, the present invention proposes the method application, specifically based on a complex industrial process evaluation of running status method of T-KPRM, and the method is applied in dense-medium separation density observing and controlling process, obtains good evaluation effect.The method, by accurately extracting the information of each variable of field data, is set up the evaluation model of historical process data, is then utilized three interpretational criterias to carry out the on-line operation state evaluation of density observing and controlling process.Finally by evaluation result, carry out the identification of running status non-optimal factor, provide decision-making foundation for operating personnel adjust production strategy in real time, thus improve production efficiency and the economic benefit of enterprise.

Claims (2)

1. based on a complex industrial process evaluation of running status method of T-KPRM, it is characterized in that, suppose that process data matrix is X ∈ R ^{n × J}, N is sample number, and J is process variable number; Exporting data matrix is Y ∈ R ⁿ, comprise output procedure variable; Concrete steps then based on the complex industrial process evaluation of running status method of T-KPRM are as follows:

Step one, zero-mean and unit variance process are carried out to each row of input data matrix X; In like manner, also standardization is carried out to output data matrix Y;

Step 2, by input data matrix X through Nonlinear Mapping Φ: x _i∈ R ⁿ→ Φ (x _i) ∈ F projects to high-dimensional feature space F, and calculate nuclear matrix K:K=Φ in F space ^tΦ;

Step 3, standardization is carried out to nuclear matrix K;

Step 4, PRM algorithm is run to input nucleus matrix K and output matrix Y, specific as follows:

A1, to establish be the lever weights of i-th sample data, can be represented by following formula:

$w_i^x=f(\frac{\left|\right|t_i-{\mathrm{med}}_{L1}\left(T\right)\left|\right|}{{\mathrm{med}}_i\left|\right|t_i-{\mathrm{med}}_{L1}\left(T\right)\left|\right|},c),i1,2,\mathrm{...}n---\left(1\right),$

And $f(z,c)=\frac{1}{{(1+|z/c\left|\right)}^2}---\left(2\right),$

Wherein, || || represent Euclidean distance, med represents median, med _l1represent L1 median, t _ibe the PLS score of i-th sample data, c is constant;

If the residual error weights of i-th sample data can be defined by following formula:

Wherein, r _irepresent the residual error between i-th sample data predicted value and actual value, represent the robust size estimation of residual error, can be calculated by following formula: $\tilde{r}={\mathrm{med}}_i|r_i-{\mathrm{med}}_j\left(r_j\right)|,i,j=1,2,\mathrm{...}n---\left(4\right);$

The then comprehensive weight w of i-th sample data _i, can be determined by following formula: utilize formula (1), (3), (5) initialization weights W _i;

B1, respectively input nucleus matrix K and output matrix Y to be weighted, after weighting, to obtain input nucleus matrix K _wwith output matrix Y _w, then PLS1 regression model is set up to the inputoutput data after weighting, revises its score vector simultaneously;

C1, calculate the residual error r of each data sample _i, utilize formula (1), (3), (5) to upgrade sample weights W _i;

If the relative difference of the double q of calculating in d1 front and back is less than threshold value, then enters step 5, otherwise return b1;

Step 5, now input data matrix become K _w, export data matrix and become Y _w, from output matrix Y _wthe middle u extracting convergence _i, make i=1, K _wi=K _w, Y _wi=Y _w;

A2, make Y _wiin any one row equal u _i;

B2, calculating K _wscore vector, t _i=K _wiu _i, t _i← t _i/ || t _i||;

c2、 $q_i=Y_{Wi}^T{t}_i;$

D2, calculating Y _wiscore vector u _i=Y _wiq _i, u _i← u _i/ || u _i||;

E2, judge u _iwhether restrain, if convergence, proceed to step 6, otherwise return a2;

Step 6, calculating K _wiload matrix:

Step 7, extract whole pivot, calculate input data matrix K _wscore matrix T, input data matrix K _wload matrix P, export data matrix Y _wscore matrix U and export data matrix Y _wload matrix Q, specific as follows:

A3, order $K_{Wi+1}=K_{Wi}-t_i{p}_i^T;Y_{Wi+1}=Y_{Wi}-u_i{q}_i^T;$

B3, make i=i+1, repeat step 5, six until extract A pivot, pivot number A can be determined by cross-validation method;

c3、T＝[t ₁,…,t _A]，P＝[p ₁,…,p _A]，U＝[u ₁,…,u _A]，Q＝[q ₁,…,q _A]；

Step 8, K _w=TP ^t+ E, Y _w=UQ ^t+ F;

Step 9, to pivot TP ^trun PCA algorithm:

Step 10, PCA algorithm is run to residual error E:

If exporting data matrix Y is single output variable, then the expression formula of T-KPRM model is as follows:

$\{\begin{array}{c}{K}_W=t_y{p}_y^T+T_o{P}_o^T+T_r{P}_r^T+E_r\\ y_W=t_y+F\end{array}---\left(6\right)$

Above-mentioned score vector t _ythat represent is in T and output y _wdirectly related part, namely sets up the most critical variable drawn required for evaluation model, can be used for the foundation of running status evaluated off-line model; T _orepresent in T with output y _worthogonal part; ; T _rit is the part having larger variance in residual error E; E _rfinal residual error, namely noise;

When introducing new sample k _new, its score matrix will be drawn by following formula:

$t_y=y_W^{T}T{\left(U^T{K}_{W}T\right)}^{-1}{U}^T{k}_{new}=g_y{k}_{new}---\left(7\right),$

Wherein, k _newnew samples x _newkernel function, can be calculated by following formula:

k _new＝Φ(X)Φ(x _new)＝[k(x ₁,x _new),…k(x _n,x _new)] ^T(8)，

To k _newcarry out equalization can obtain: ${\stackrel{\&OverBar;}{k}}_{new}=k_{new}-K{1}_t-1_n{k}_{new}+1_{n}K{1}_t---\left(9\right),$

Wherein, 1 _t=1/n [11...1] ^t∈ R ⁿ.

2. the application of a kind of complex industrial process evaluation of running status method based on T-KPRM according to claim 1, it is characterized in that, a kind of application of the complex industrial process evaluation of running status method based on T-KPRM refers to the identification utilizing core inclined robust latent variable technology to set up the evaluated off-line model of running status, the on-line evaluation of complex industrial process running status and non-optimal factor, specifically first utilize core inclined robust latent variable technology to set up evaluated off-line model, the evaluated off-line model score vector of corresponding running status grade is wherein, c is the number of running status grade; Then, introduce sliding window technique, calculate the similarity between online data window and corresponding opinion rating, utilize similarity to carry out the on-line evaluation of industrial process running status; Euclidean distance between recycling slip data window and optimum opinion rating, calculates the contribution rate of relevant variable, identifies the non-optimal factor of running status.