CN104537260A - Dynamic soft measurement method and system based on slow characteristic regression - Google Patents

Abstract

The invention relates to a dynamic soft measurement method and system based on slow characteristic regression. The method includes the steps of reading a slow characteristic model and a linear regression model, reading in online measurement values of operating variables, inputting the online measurement values into the slow characteristic model, and obtaining an online real-time estimated value of a slow characteristic vector s(t); inputting a vector s<1:M>(t) composed of the first M slow characteristics into the linear regression model, and obtaining an instant estimated value (please see the specification) of the leading variable y(t) difficult to measure; writing in and displaying the estimated value (please see the specification) of the y(t). By means of the dynamic soft measurement method and system in the technical scheme, the dynamic characteristics of a process can be effectively mined, the estimation accuracy is higher, the more excellent control performance is achieved when closed-loop feedback control is implemented, the prediction accuracy of a current model can be visually and effectively judged according to the slow change degrees of the hidden characteristics when field implementation is carried out, and the information in large-scale process data can be effectively used.

Description

The dynamic soft-measuring method returned based on slow feature and system

Technical field

The present invention relates to dynamic analysis technology field, in particular to a kind of dynamic soft-measuring method of returning based on slow feature and a kind of dynamic soft measuring system returned based on slow feature.

Background technology

In industrial processes, the index that the quality of product is usually important with some is relevant, the doing of such as diesel oil, polyacrylic melting index, the pH value etc. in chemical reactor.Control these key indexs to reach preset value and usually become target in production.But, these indexs cannot carry out on-line measurement usually, and the lab analysis chemical examination be merely able to by reaching several hours obtains, and the analysis data in laboratory postpones too large, cannot be used as the feedback signal of control loop, the control overflow of high-quality is not being met; In addition, the online analytical instrument of some index is very expensive, and later maintenance costly.Based on above two reasons, as a kind of emerging online variable estimation technique---hard measurement (soft instrument) modeling is arisen at the historic moment, in nearly 30 years, obtain significant progress, and be widely used in the actual industrial such as chemical process, bio-pharmaceuticals, iron and steel forging.Flexible measurement method is by setting up the mathematical relation Modling model between secondary variable and quality variable, and the real-time prediction realizing surveying difficulty variable exports.

Chemical process important feature is that it exists obvious behavioral characteristics, that is, the quality variable of current time is not only subject to the impact of current time process variable, and all there is certain relation with the process variable in settling time.In order to consider dynamic characteristic of the course, scholars, by adding the historical data of certain length in process input, are improved traditional soft-measuring modeling method, and are achieved certain achievement.But because the sample frequency of process variable is far away higher than quality variable, the sample size that can be used in soft sensor modeling is limited; When considering the historical data of certain length, the input dimension of soft-sensing model can be multiplied, and over-fitting problem very easily occurs model.In addition, when describing process dynamics, the implicit structure of latent variable model (as offset minimum binary etc.) lacks clear and definite physical significance, makes the interpretation of model poor.Therefore, according to the model setting up suitable description dynamic characteristic of the course, strengthen the precision of prediction of soft-sensing model further, there is very important practical significance.

Summary of the invention

Technical matters to be solved by this invention is, how to solve On-line sampling system problem difficulty important in production run being surveyed to variable.

For this purpose, the present invention proposes a kind of dynamic soft-measuring method returned based on slow feature, comprising: S1, read slow characteristic model and linear regression model (LRM), described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, then the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c,

Wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^m, c ∈ R is respectively coefficient vector and the constant term of linear regression model (LRM);

S2, reads in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t);

S3, by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

S4, by the estimated value of y (t) write and show.

Preferably, the process of establishing of described slow characteristic model comprises: S10, builds a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, arranges the restriction on the parameters scope of structural relation,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

S11, collects relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and forms sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem.

S12, to C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

S13, according to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

S14, definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

S15, by matrix Z compute matrix

$\stackrel{\&CenterDot;}{Z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix; Without loss of generality, suppose that the element on Ω diagonal line arranges according to order from small to large.

S16, compute matrix W=PQ.

Preferably, the process of establishing of described linear regression model (LRM) comprises:

S17, collects relevant easy survey auxiliary variable and surveys leading variable data form sample set C with difficult from the database and off-line analysis data storehouse of online acquisition _s,

$,C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data;

S18, calculates C according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

S19, for each k ∈ 1 ..., m} carries out cross validation, utilizes s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& y\left(T_{N_s}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression.

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

S20, according to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$

The invention allows for a kind of dynamic soft measuring system returned based on slow feature, comprising: reading unit, read slow characteristic model and linear regression model (LRM), described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, then the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c,

Wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^mcoefficient vector and the constant term of linear regression model (LRM) is respectively with c ∈ R;

Model treatment unit, for reading in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t), and by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

Display unit, for the estimated value by y (t) write and show.

Preferably, also comprise: slowly characteristic model sets up unit, for building a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, the restriction on the parameters scope of structural relation be set,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

Collect relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and form sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem.

To C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

According to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

Definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

By matrix Z compute matrix

$\stackrel{\&CenterDot;}{Z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix, and the element on matrix Ω diagonal line arranges according to order from small to large,

Compute matrix W=PQ.

Preferably, also comprise: linear regression model unit, survey leading variable data for collecting relevant easy survey auxiliary variable from the database and off-line analysis data storehouse of online acquisition form sample set C with difficult _s,

$C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data, the sampling period is generally longer, and usually has atypical characteristic;

C is calculated according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

For each k ∈ 1 ..., m} carries out cross validation, utilizes s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& y\left(T_{N_s}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression,

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

According to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$

By technical scheme of the present invention, at least following technique effect can be realized:

1) model that the present invention proposes is applicable to the obvious soft sensor modeling problem of behavioral characteristics, can the behavioral characteristics of mining process effectively.Owing to considering the history input in the dynamic transition process time, existing partial dynamic soft-sensing model input dimension is too high, and then result in over-fitting problem.The model that the present invention proposes efficiently solves the problems referred to above, and compared with conventional dynamic soft-sensing model (such as dynamic offset minimum binary), estimated accuracy is higher;

2) because Quality Forecasting value is obtained by the Feature Combination slowly changed, the forecast output of model itself has certain slickness, more identical with the actual characteristic of process, also has more excellent control performance when implementing close-loop feedback control;

3) the slow intensity of variation of this model to each hidden feature quantizes, there is better physics interpretation, when thus can implement at the scene, the slow intensity of variation according to each hidden feature carries out intuitively, effectively judging to the precision of prediction of "current" model;

4) traditional soft-measuring modeling method belongs to the category of supervised learning, and the training process of the model that the present invention proposes is semi-supervised, effectively can utilize the information in large scale process data.

Accompanying drawing explanation

Can understanding the features and advantages of the present invention clearly by reference to accompanying drawing, accompanying drawing is schematic and should not be construed as and carry out any restriction to the present invention, in the accompanying drawings:

Fig. 1 shows the schematic flow diagram of the dynamic soft-measuring method returned based on slow feature according to an embodiment of the invention;

Fig. 2 A shows the schematic flow diagram setting up slow characteristic model according to an embodiment of the invention;

Fig. 2 B shows the schematic flow diagram setting up linear regression model (LRM) according to an embodiment of the invention;

Fig. 3 shows the principle schematic of the dynamic soft-measuring method returned based on slow feature according to an embodiment of the invention;

Fig. 4 shows the schematic block diagram of the dynamic soft measuring system returned in slow feature according to an embodiment of the invention;

Fig. 5 shows host computer implementation according to an embodiment of the invention;

Fig. 6 shows according to an embodiment of the invention as d=5, the schematic diagram of front 4 slow features;

Fig. 7 shows according to an embodiment of the invention as d=5, and the estimation of difficult survey variable compares schematic diagram with actual value.

Embodiment

Be described below in detail embodiments of the invention, the example of described embodiment is shown in the drawings, and wherein same or similar label represents same or similar element or has element that is identical or similar functions from start to finish.Being exemplary below by the embodiment be described with reference to the drawings, only for explaining the present invention, and can not limitation of the present invention being interpreted as.

Those skilled in the art of the present technique are appreciated that unless expressly stated, and singulative used herein " ", " one ", " described " and " being somebody's turn to do " also can comprise plural form.Should be further understood that, the wording used in instructions of the present invention " comprises " and refers to there is described feature, integer, step, operation, element and/or assembly, but does not get rid of and exist or add other features one or more, integer, step, operation, element, assembly and/or their group.Should be appreciated that, when we claim element to be " connected " or " coupling " to another element time, it can be directly connected or coupled to other elements, or also can there is intermediary element.In addition, " connection " used herein or " coupling " can comprise wireless connections or wirelessly to couple.Wording "and/or" used herein comprises one or more whole or arbitrary unit listing item be associated and all combinations.

Those skilled in the art of the present technique are appreciated that unless otherwise defined, and all terms used herein (comprising technical term and scientific terminology), have the meaning identical with the general understanding of the those of ordinary skill in field belonging to the present invention.It should also be understood that, those terms defined in such as general dictionary, should be understood to that there is the meaning consistent with the meaning in the context of prior art, unless and by specific definitions as here, otherwise can not explain by idealized or too formal implication.

Those skilled in the art of the present technique are appreciated that, here used " terminal ", " terminal device " had both comprised the equipment of wireless signal receiver, it only possesses the equipment of the wireless signal receiver without emissive ability, comprise again the equipment receiving and launch hardware, it has and on bidirectional communication link, can perform the reception of two-way communication and launch the equipment of hardware.This equipment can comprise: honeycomb or other communication facilitiess, its honeycomb or other communication facilities of having single line display or multi-line display or not having multi-line display; PCS (PersonalCommunications Service, PCS Personal Communications System), it can combine voice, data processing, fax and/or its communication ability; PDA (Personal Digital Assistant, personal digital assistant), it can comprise radio frequency receiver, pager, the Internet/intranet access, web browser, notepad, calendar and/or GPS (Global Positioning System, GPS) receiver; Conventional laptop and/or palmtop computer or other equipment, it has and/or comprises the conventional laptop of radio frequency receiver and/or palmtop computer or other equipment.Here used " terminal ", " terminal device " can be portable, can transport, be arranged in the vehicles (aviation, sea-freight and/or land), or be suitable for and/or be configured at local runtime, and/or with distribution form, any other position operating in the earth and/or space is run.Here used " terminal ", " terminal device " can also be communication terminal, access terminals, music/video playback terminal, can be such as PDA, MID (Mobile InternetDevice, mobile internet device) and/or there is the mobile phone of music/video playing function, also can be the equipment such as intelligent television, Set Top Box.

Those skilled in the art of the present technique are appreciated that, the concepts such as server used here, high in the clouds, remote network devices, have effects equivalent, it includes but not limited to the cloud that computing machine, network host, single network server, multiple webserver collection or multiple server are formed.At this, cloud is formed by based on a large amount of computing machine of cloud computing (Cloud Computing) or the webserver, and wherein, cloud computing is the one of Distributed Calculation, the super virtual machine be made up of a group loosely-coupled computing machine collection.In embodiments of the invention, realize communicating by any communication mode between remote network devices, terminal device with WNS server, include but not limited to, the mobile communication based on 3GPP, LTE, WIMAX, the computer network communication based on TCP/IP, udp protocol and the low coverage wireless transmission method based on bluetooth, Infrared Transmission standard.

Those skilled in the art are to be understood that, " application ", " application program ", " application software " alleged by the present invention and the concept of similar statement, be those skilled in the art known same concept, refer to and be suitable for by the instruction of series of computation machine and the organic structure of related data resource the computer software that electronics runs.Unless specified, this name itself, not by programming language kind, rank, also not limited by the operating system of its operation of relying or platform.In the nature of things, this genus also not limited by any type of terminal.

As shown in Figure 1, according to an embodiment of the invention based on the dynamic soft-measuring method that slow feature returns, comprising: S1, read slow characteristic model and linear regression model (LRM), described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, then the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c,

Wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^mcoefficient vector and the constant term of linear regression model (LRM) is respectively with c ∈ R;

S2, reads in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t);

S3, by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

S4, by the estimated value of y (t) write and show.

As shown in Figure 2 A, preferably, the process of establishing of described slow characteristic model comprises:

S10, builds a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, arranges the restriction on the parameters scope of structural relation,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

S11, collects relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and forms sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem.

S12, to C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

S13, according to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

S14, definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

S15, by matrix Z compute matrix

$\stackrel{\&CenterDot;}{Z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix; Without loss of generality, suppose that the element on Ω diagonal line arranges according to order from small to large.

S16, compute matrix W=PQ.

As shown in Figure 2 B, preferably, the process of establishing of described linear regression model (LRM) comprises:

S17, collects relevant easy survey auxiliary variable and surveys leading variable data form sample set C with difficult from the database and off-line analysis data storehouse of online acquisition _s,

$C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data, preferably, for the data-driven method of problematic features, system identifying method generally cannot be utilized to solve, the sampling period therefore in the present invention is general longer, and usually has atypical characteristic,

S18, calculates C according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

S19, for each k ∈ 1 ..., m} carries out cross validation, can select optimum M value, then utilize s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& y\left(T_{N_s}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression.

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

S20, according to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$

Said method realize principle as shown in Figure 3.

As shown in Figure 4, according to an embodiment of the invention based on the dynamic soft measuring system 10 that slow feature returns, comprising: reading unit 11, read slow characteristic model and linear regression model (LRM), described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, then the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c,

Wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^mcoefficient vector and the constant term of linear regression model (LRM) is respectively with c ∈ R;

Model treatment unit 12, for reading in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t), and by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

Display unit 13, for the estimated value by y (t) write and show.

Preferably, also comprise:

Slow characteristic model sets up unit 14, for building a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, arranges the restriction on the parameters scope of structural relation,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$x\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_2(t-\mathrm{d\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ \&CenterDot;\\ \&CenterDot;\\ \&CenterDot;\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

Collect relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and form sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem;

To C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

According to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

Definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

By matrix Z compute matrix

$\stackrel{\&CenterDot;}{Z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix, and the element on matrix Ω diagonal line arranges according to order from small to large, last compute matrix W=PQ.

Preferably, also comprise: linear regression model unit 15, from the database and off-line analysis data storehouse of online acquisition, collect relevant easy survey auxiliary variable survey leading variable data form sample set C with difficult _s,

$C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data, the sampling period is generally longer, and usually has atypical characteristic;

C is calculated according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

For each k ∈ 1 ..., m} carries out cross validation, utilizes s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& y\left(T_{N_s}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression.

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

According to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& \&CenterDot;\&CenterDot;\&CenterDot;& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& \&CenterDot;\&CenterDot;\&CenterDot;& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$

As shown in Figure 5, supervisory system can be realized by host computer in industry spot according to an embodiment of the invention.Data sampling and processing in the present invention and Optimization Modeling, on-line prediction can be realized by host computer.Control program, by real-time data base or by OPC (OLE forProcess Control) mode retrieve processed data, shows at host computer after main data processed result has calculated.

Certainly, technical scheme disclosed in this invention can also adopt terminal or server to be that carrier realizes.Concrete selection can be determined according to actual needs.

Below by way of instantiation, dynamic soft measuring process of the present invention is described:

For close to actual production process, the selected Tennessee-Yi Siman process (TE Process) proposed by Eastman Chemical Company.Tennessee-Yi Siman process is that process control provides an extraordinary research platform, plays the role of benchmark problem in the problems such as process monitoring, hard measurement, fault diagnosis.Whole process is made up of reactor, condenser, compressor, gas-liquid separator and stripping tower.The reactant A of four kinds of gaseous states, C, D, E and inert gas B deliver in reactor as charging, finally obtain two kinds of product G and H, and secondary product F.The homepage that this data set can teach Richard D.Braatz from Massachusetts Institute of Technology (MIT) is downloaded.

In the present embodiment, have 33 process variable, be chosen as XMV (1-11) and XMEAS (1-22), wherein XMV (1-11) is manipulated variable, and XMEAS (1-22) is measurand.Difficult survey variables choice is the content of reactant C in final products.The sampling period of process variable is 3 minutes, and the difficult sampling period of surveying variable is 6 minutes.The raw training sample 500 of common property, test sample book 960.Error measurement index is that square error is weighed

$\mathrm{RMSE}=\sqrt{\frac{1}{960}\underset{i=1}{\overset{960}{\mathrm{\&Sigma;}}}{[\hat{y}\left(T_i\right)-y\left(T_i\right)]}^2}.$

And the related coefficient of predicted value and actual value (r value).The span of length of history data d be 0,1 ..., 7}.

Set up slow characteristic model and linear regression model (LRM) according to above-mentioned model process of establishing, based on the model established, implement on-line measurement.

Steps A 1: read slow characteristic model and linear regression model (LRM) respectively.

Steps A 2: the on-line measurement value of reading in performance variable in DCS system (specifically can be realized by terminal or server), and be input in slow characteristic model, try to achieve the online estimated value in real time of slow proper vector s (t);

Steps A 3: by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t);

Steps A 4: the estimated value of y (t) is written in DCS system and shows, for operative employee's reference, instruct practical operation, or for controlling and optimizing.In this example, the method returned based on slow feature and traditional dynamic offset minimum binary (Dynamic PLS) method are contrasted, result is as shown in table 1:

Table 1

Fig. 6 and Fig. 7 illustrates the partial data in table comparatively intuitively.Test result shows that this model is compared with traditional dynamic soft measuring modeling method, and predicated error is less, the degree of correlation of predicted value and actual value is higher, has higher precision of prediction.As latent variable model, when length of history data increases, model should excavate the artificial mixed forests information more enriched from the time series data of process; But no matter length of history data is increased to how many, and conventional dynamic offset minimum binary can only obtain 1 hidden variable, and there is any discrepancy with practical experience, there is obvious deficiency.As a comparison, the flexible measurement method returned based on slow feature can excavate out more how significant hidden variable, and along with the increase of length of history data, the number of hidden variable increases thereupon, more meets the objective characteristic of process.

In sum, the model that the present invention proposes is applicable to the obvious soft sensor modeling problem of behavioral characteristics, can the behavioral characteristics of mining process effectively.Owing to considering the history input in the dynamic transition process time, existing partial dynamic soft-sensing model input dimension is too high, and then result in over-fitting problem.The model that the present invention proposes efficiently solves the problems referred to above, and compared with conventional dynamic soft-sensing model (such as dynamic offset minimum binary), estimated accuracy is higher; Because Quality Forecasting value is obtained by the Feature Combination slowly changed, the forecast output of model itself has certain slickness, more identical with the actual characteristic of process, also has more excellent control performance when implementing close-loop feedback control; The slow intensity of variation of this model to each hidden feature quantizes, and has better physics interpretation, and when thus can implement at the scene, the slow intensity of variation according to each hidden feature carries out intuitively, effectively judging to the precision of prediction of "current" model; Tradition soft-measuring modeling method belongs to the category of supervised learning, and the training process of the model that the present invention proposes is semi-supervised, effectively can utilize the information in large scale process data.

It should be noted that the algorithm provided at this is intrinsic not relevant to any certain computer, virtual system or miscellaneous equipment with formula.Various general-purpose system also can with use based on together with this example.According to description above, the structure constructed required by this type systematic is apparent.In addition, the present invention is not also for any certain programmed language.It should be understood that and various programming language can be utilized to realize content of the present invention described here, and the description done language-specific is above to disclose preferred forms of the present invention.

In instructions provided herein, describe a large amount of detail.But can understand, embodiments of the invention can be put into practice when not having these details.In some instances, be not shown specifically known method, structure and technology, so that not fuzzy understanding of this description.

Similarly, be to be understood that, in order to simplify the present invention and to help to understand in various aspects of the present invention one or more, in the description above to exemplary embodiment of the present invention, each feature of the present invention is grouped together in single embodiment, figure or the description to it sometimes.But, the method and apparatus of the disclosure should be construed to the following intention of reflection: namely the present invention for required protection requires feature more more than the feature clearly recorded in each claim.Or rather, as claims reflect, all features of disclosed single embodiment before inventive aspect is to be less than.Therefore, the claims following embodiment are incorporated to this embodiment thus clearly, and wherein each claim itself is as independent embodiment of the present invention.

Those skilled in the art are appreciated that and adaptively can change the module in the equipment in embodiment and they are arranged in one or more equipment different from this embodiment.Module in embodiment or unit or assembly can be combined into a module or unit or assembly, and multiple submodule or subelement or sub-component can be put them in addition.Except at least some in such feature and/or process or unit be mutually repel except, any combination can be adopted to combine all processes of all features disclosed in this instructions (comprising adjoint claim, summary and accompanying drawing) and so disclosed any method or equipment or unit.Unless expressly stated otherwise, each feature disclosed in this instructions (comprising adjoint claim, summary and accompanying drawing) can by providing identical, alternative features that is equivalent or similar object replaces.

In addition, those skilled in the art can understand, although embodiments more described herein to comprise in other embodiment some included feature instead of further feature, the combination of the feature of different embodiment means and to be within scope of the present invention and to form different embodiments.

All parts embodiment of the present invention with hardware implementing, or can realize with the software module run on one or more processor, or realizes with their combination.It will be understood by those of skill in the art that the some or all functions that microprocessor or digital signal processor (DSP) can be used in practice to realize according to the some or all parts in the web portal security checkout equipment of the embodiment of the present invention.The present invention can also be embodied as part or all equipment for performing method as described herein or device program (such as, computer program and computer program).Realizing program of the present invention and can store on a computer-readable medium like this, or the form of one or more signal can be had.Such signal can be downloaded from internet website and obtain, or provides on carrier signal, or provides with any other form.

The above is only some embodiments of the present invention; it should be pointed out that for those skilled in the art, under the premise without departing from the principles of the invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.

Claims (6)

1. based on the dynamic soft-measuring method that slow feature returns, it is characterized in that, comprising:

S1, reads slow characteristic model and linear regression model (LRM), and described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ .\\ .\\ .\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, then the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_2(t-\mathrm{d\&Delta;t})\\ .\\ .\\ .\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c, wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^mcoefficient vector and the constant term of linear regression model (LRM) is respectively with c ∈ R;

S2, reads in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t);

S3, by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

S4, by the estimated value of y (t) write and show.

2. according to claim 1 based on the dynamic soft-measuring method that slow feature returns, it is characterized in that, the process of establishing of described slow characteristic model comprises:

S10, builds a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, arranges the restriction on the parameters scope of structural relation,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ .\\ .\\ .\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_2(t-\mathrm{d\&Delta;t})\\ .\\ .\\ .\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

S11, collects relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and forms sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem,

S12, to C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

S13, according to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

S14, definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

S15, by matrix Z compute matrix

$\stackrel{\&CenterDot;}{z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix, and the element on matrix Ω diagonal line is according to order arrangement from small to large,

S16, compute matrix W=PQ.

3. according to claim 1 based on the dynamic soft-measuring method that slow feature returns, it is characterized in that, the process of establishing of described linear regression model (LRM) comprises:

S17, collects relevant easy survey auxiliary variable and surveys leading variable data form sample set C with difficult from the database and off-line analysis data storehouse of online acquisition _s,

$C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data;

S18, calculates C according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

S19, for each k ∈ 1 ..., m} carries out cross validation, utilizes s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& ...& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& ...& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& ...& y\left(T_{N_a}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression,

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

S20, according to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& ...& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& ...& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$

4., based on the dynamic soft measuring system that slow feature returns, it is characterized in that, comprising:

Reading unit, reads slow characteristic model and linear regression model (LRM), and described slow characteristic model is

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ .\\ .\\ .\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right),$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and described input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of described input vector u (t) is

$u\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_2(t-\mathrm{d\&Delta;t})\\ .\\ .\\ .\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and dimension m meets relational expression m=n (d+1);

Described linear regression model (LRM) is y (t)=b ^ts _1:M(t)+c, wherein M is the number for the slow feature predicted,

S _1:M(t)=[s ₁(t) s ₂(t) ... s _m(t)] ^tfor the vector be made up of front M slow feature, b ∈ R ^mcoefficient vector and the constant term of linear regression model (LRM) is respectively with c ∈ R;

Model treatment unit, for reading in the on-line measurement value of performance variable, and is input in described slow characteristic model, tries to achieve the online estimated value in real time of slow proper vector s (t), and by the vectorial s of front M slow feature composition _1:Mt () is input in linear regression model (LRM), obtain the difficult instantaneous estimation value surveying leading variable y (t)

Display unit, for the estimated value by y (t) write and show.

5. hard measurement system according to claim 4, is characterized in that, also comprise:

Slow characteristic model sets up unit, for building a difficult structural relation surveyed between leading variable and multiple easy survey auxiliary variable, arranges the restriction on the parameters scope of structural relation,

Described structural relation is:

Output quantity y represents difficult survey leading variable,

Input quantity x (t)=[x ₁(t), x ₂(t) ..., x _n(t)] ^tthe easy survey auxiliary variable that representative is relevant,

Structural relation is in series by Dynamic mode and linear-static link, and the output of Dynamic mode is one group of slow feature:

$s\left(t\right)=\left[\begin{array}{c}{s}_1\left(t\right)\\ s_2\left(t\right)\\ .\\ .\\ .\\ s_m\left(t\right)\end{array}\right]=\mathrm{Wu}\left(t\right)$

Wherein, m is the dimension of input vector, and the m that u (t) is sampling instant discretize ties up input vector, and the m that s (t) is sampling instant discretize ties up slow proper vector; Matrix of coefficients W is m rank square formations, and input vector u (t) comprises the historical data of some, easily surveys auxiliary variable x (t)=[x when there being n ₁(t), x ₂(t) ..., x _n(t)] ^ttime, the structure of input vector u (t) is:

$x\left(t\right)=\left[\begin{array}{c}{x}_1\left(t\right)\\ x_1(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_1(t-\mathrm{d\&Delta;t})\\ x_2\left(t\right)\\ x_2(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_2(t-\mathrm{d\&Delta;t})\\ .\\ .\\ .\\ x_n\left(t\right)\\ x_n(t-\mathrm{\&Delta;t})\\ .\\ .\\ .\\ x_n(t-\mathrm{d\&Delta;t})\end{array}\right]\&Element;R^{n(d+1)}$

Wherein, Δ t is the sampling interval of input variable, and d is the length of the historical data that input vector comprises, and the dimension m of input vector meets relational expression m=n (d+1), according to priori data, and the estimation of calculated transient time d;

Collect relevant easy survey auxiliary variable from the database of online acquisition by the order that the sampling time increases progressively and form sample set C _u:

C _u＝{u(t),u(t+Δt),…,u(t+(N _u-1)Δt)}

Number of samples is N _u, and sampling period Δ t meets Shannon's sampling theorem;

To C _uin input amendment data carry out standardization, make input data have zero-mean and unit variance in each dimension;

According to C _uin composition of sample matrix U,

$U={[u\left(t\right),u(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,u(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u}$

To matrix U U ^tcarry out Eigenvalues Decomposition, UU ^t=V Λ V ^t,

Wherein, V is orthogonal matrix, and Λ is diagonal matrix;

Definition matrix Q=Λ ^-1/2v ^t, by matrix Q, u (t) is converted, obtains intermediate variable z (t)=Qu (t),

And matrix $Z={[z\left(t\right),z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;N_u};$

By matrix Z compute matrix

$\stackrel{\&CenterDot;}{z}=\frac{1}{\mathrm{\&Delta;t}}{[z(t+\mathrm{\&Delta;t})-z\left(t\right),z(t+2\mathrm{\&Delta;t})-z(t+\mathrm{\&Delta;t}),\&CenterDot;\&CenterDot;\&CenterDot;,z(t+(N_u-1)\mathrm{\&Delta;t})-z(t+(N_u-2)\mathrm{\&Delta;t})]}^T\&Element;R^{m\&times;(N_u-1)}$

To matrix carry out Eigenvalues Decomposition

Wherein, P is orthogonal matrix, and Ω is diagonal matrix, and the element on matrix Ω diagonal line arranges according to order from small to large,

Compute matrix W=PQ.

6. hard measurement system according to claim 4, is characterized in that, also comprise:

Linear regression model unit, collects relevant easy survey auxiliary variable and surveys leading variable data form sample set C with difficult from the database and off-line analysis data storehouse of online acquisition _s,

$C_s=\{(u\left(T_1\right),y\left(T_1\right)),(u\left(T_2\right),y\left(T_2\right)),\&CenterDot;\&CenterDot;\&CenterDot;(u\left(T_{N_s}\right),y\left(T_{N_s}\right))\}$

Number of samples is N _s, T _ithe sampling instant of difficult survey leading variable i-th off-line sampled data, y (T _i) (i=1 ..., N _s) difficult survey leading variable i-th off-line sampled data;

C is calculated according to matrix W _sin the slow feature of each sample:

s(T _i)＝Wx(T _i),1≤i≤N _s；

For each k ∈ 1 ..., m} carries out cross validation, utilizes s (T _i) front k element s _1:k(T _i) ∈ R ^k, structural matrix S _1:k:

$S_{1:k}=\left[\begin{array}{cccc}{s}_{1:k}\left(T_1\right)& s_{1:k}\left(T_2\right)& ...& s_{1:k}\left(T_{N_s}\right)\\ 1& 1& ...& 1\end{array}\right]\&Element;R^{(k+1)\&times;N_s},$

And matrix $Y={\left[\begin{array}{cccc}y\left(T_1\right)& y\left(T_2\right)& ...& y\left(T_{N_a}\right)\end{array}\right]}^T,$

Default modeling sample in S, Y}, fetching portion sample modling model,

$\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right]={\left(S_{1:k}^{\mathrm{train}}{\left(S_{1:k}^{\mathrm{train}}\right)}^T\right)}^{-1}{S}_{1:k}^{\mathrm{train}}{Y}_{1:k}^{\mathrm{train}},$

Wherein with be respectively coefficient vector and the constant term of linear regression,

By remaining sample forecast:

${\hat{Y}}^{\mathrm{vld}}={\left(S_{1:k}^{\mathrm{vld}}\right)}^T\left[\begin{array}{c}{b}_{1:k}^{\mathrm{train}}\\ c_{1:k}^{\mathrm{train}}\end{array}\right],$

The quadratic sum of record prediction error until all samples all by and only forecast once, specify k for one, obtain the prediction error quadratic sum of all samples of its correspondence, be defined as PRESS _k,

According to the prediction error quadratic sum of each k, determine the optimal value of M:

$M=\underset{k}{\arg }\min \left\{{\mathrm{PRESS}}_k\right\};$

According to s (T _i) front M element s _1:M(T _i) ∈ R ^m, structural matrix S:

$S=\left[\begin{array}{cccc}{s}_{1:M}\left(T_1\right)& s_{1:M}\left(T_2\right)& ...& s_{1:M}\left(T_{N_s}\right)\\ 1& 1& ...& 1\end{array}\right]\&Element;R^{(M+1)\&times;N_s},$

Calculate coefficient vector and the constant term of described linear regression model (LRM)

$\left[\begin{array}{c}b\\ c\end{array}\right]={\left({\mathrm{SS}}^T\right)}^{-1}\mathrm{SY}.$