CN103150482B - A kind of method determining the baking silk effects of process parameters factor based on PLS - Google Patents

Abstract

The invention discloses a kind of method of determining to dry the silk effects of process parameters factor based on PLS, by target variable and the independent variable of process data in selected baking silk technological process, gather the historical data of target variable and independent variable; Adopt Euclidean distance method to screen historical data, reject the abnormal data in historical data, obtain sample data; Adopt least square method to set up the PLS model of independent variable x and target variable y, calculate the size of the factor of influence of the independent variable x obtaining each target variable y of impact according to model.The method is simple, be easy to realize, and application this method substitutes artificial experience and judges, obtains the size of factor of influence fast and accurately.

Description

A kind of method determining the baking silk effects of process parameters factor based on PLS

Technical field

The present invention relates to a kind of method of determining to dry the silk effects of process parameters factor based on PLS.

Background technology

Baking tabacco scrap technique is requisite link in Cigarette processing Primary Processing, and drying silk technological process affects larger on the physical qualities of cigarette product and aesthetic quality.The parameter that silk technique is dried in impact has more than 10, and present stage can adjust state parameter on supervisory control comuter.But occur that (device parameter detected value occurs abnormal extremely when producing, supplied materials offset setting etc.), cannot according to the factor of influence size of drying parameter in silk technique, find accurately, rapidly and the root of unusual condition occurs, and equivalent compensation can only be made by automatic control system and process experiences does artificial adjustment.Therefore, be badly in need of proposing a kind of method and determine to dry in silk technique each parameter in whole production run to the size of the factor of influence of other parameters.

Summary of the invention

The invention provides a kind of method of determining to dry the silk effects of process parameters factor based on PLS, its object is to, overcome in prior art the size cannot determining to dry the silk effects of process parameters factor.

The technical solution used in the present invention is as follows:

Determine a method of drying the silk technogenic influence factor based on PLS, comprise the following steps:

Step 1: independent variable x and the target variable y of silk technique are dried in setting, independent variable x chooses from HT vapor pressure, HT steam flow, entrance moisture, drum inlet hot blast temperature, drum inlet hot blast air quantity, barrel temperature, barrel vapor pressure, humidity discharging throttle opening and discharge cover negative pressure, target variable y dries silk water percentage from discharging, discharging is dried silk temperature and humidity discharging humidity and chosen, gather the historical data of independent variable x and target variable y, wherein the employing frequency of historical data is T;

Step 2: adopt Euclidean distance method to screen historical data, reject the abnormal data in historical data, obtain n group sample data, sample data forms target variable matrix Y _njwith independent variable matrix X _ni, target variable Y has j, and independent variable X has i;

Described abnormal data refers to the baking silk process data drying and gather when silk process equipment breaks down;

Calculate the correlation coefficient matrix between independent variable x and independent variable x, between target variable y and target variable y and between independent variable x and target variable y drying silk technique;

By this correlation matrix, we find to dry in silk technique and there is serious multiple correlation between each independent variable, show can not there is simple linear relationship between independent variable x and target variable y, we determine to select PLS model to analyze the interact relation of independent variable to target variable under these conditions simultaneously.

Step 3: the sample data after the screening obtained according to step 2, sets up independent variable X _niwith target variable Y _njpartial least square method PLS model, calculate the size of factor of influence of each independent variable x of each target variable y of impact according to PLS model, its specific operation process is as follows:

1) the PLS component number in partial least square method PLS and cross validation test method determination sample data is utilized to be h;

PLS and principal component analysis (PCA) (PCA) its difference are for describing in variable Y while the factor also for describing variable X.In order to realize this point, be mathematically the factor of removing compute matrix X with the row of matrix Y, meanwhile, the factor of matrix Y then goes prediction by the row of matrix X, in the hope of take into account X and Y all matching obtain reasonable load P LS composition.

2) adopt the iterative algorithm in partial least square method, be obtained from matrix of variables X _niweight matrix W, objective matrix Y _njload factor matrix Q ^twith independent variable matrix X _niloading matrix P ^t; Independent variable matrix X _niweight matrix W, objective matrix Y _njload factor matrix Q ^twith independent variable matrix X _niloading matrix P ^tbe calculated as prior art, weight matrix W is h capable i row, load factor matrix Q ^tthe capable j row of h, loading matrix P ^tthe capable i row of h;

$W=\left[\begin{array}{cccc}{w}_{11}& ...& & w_{1i}\\ w_{21}& ...& & w_{2i}\\ ...& ...& & ...\\ ...& ...& & ...\\ w_{h1}& ...& & w_{\mathrm{hi}}\end{array}\right]$

$Q^T=\left[\begin{array}{cccc}{a}_{11}& ...& & a_{1j}\\ a_{21}& ...& & a_{2j}\\ ...& ...& & ...\\ ...& ...& & ...\\ a_{h1}& ...& & a_{\mathrm{hj}}\end{array}\right],P^T=\left[\begin{array}{cccc}{c}_{11}& ...& & c_{1i}\\ c_{21}& ...& & c_{2i}\\ ...& ...& & ...\\ ...& ...& & ...\\ c_{h1}& ...& & c_{\mathrm{hi}}\end{array}\right]$

3) that determines independent variable matrix X obtains resolute t _hwith objective matrix Y resolute u _h:

Independent variable matrix X obtains resolute t _hfor:

t _h=c _h1x ₁+c _h2x ₂+…+c _hix _i

Objective matrix Y obtains resolute u _hfor:

u _h=a _h1y ₁+…+a _hjy _j

Wherein, h is the component number of sample data, and i is independent variable number, and j is target variable number;

4) be normalized calculating to the weight matrix W of independent variable matrix X, obtaining two inside correlation matrixes obtained between resolute is V:

$V=\left[\begin{array}{cccc}{v}_1& 0& ...& 0\\ 0& ...& ...& ...\\ ...& ...& ...& ...\\ ...& ...& ...& 0\\ 0& ...& 0& v_h\end{array}\right]$

Thus two internal relations formulas obtained between resolute are expressed as:

U ₁=v ₁t ₁+ e, u ₂=v ₂t ₂+ e, u ₃=v ₃t ₃+ e ... u _h=v _ht _h+ e, e are regression residuals;

5) according to B=V (P ^tv) ^-1q ^tcalculate the matrix of coefficients B in Y=XB:

$B=\left[\begin{array}{cccc}{b}_{11}& ...& & b_{1i}\\ b_{21}& ...& & b_{2i}\\ ...& ...& & ...\\ ...& ...& & ...\\ b_{j1}& ...& & b_{\mathrm{ji}}\end{array}\right]$

Thus the relation finally obtained between independent variable x and target variable y is as follows:

y ₁=b ₁₁x ₁+b ₁₂x ₂+b ₁₃x ₃+……+b _1ix _i

y ₂=b ₂₁x ₁+b ₂₂x ₂+b ₂₃x ₃+……+b _2ix _i

……

y _j=b _j1x ₁+b _j2x ₂+b _j3x ₃+……+b _jix _i

Step 4: extract the factor of influence of independent variable as corresponding target variable that independent variable absolute coefficient in above-mentioned formula is greater than 1/i, i is the number of independent variable X.

If the absolute value of independent variable coefficient is greater than 1/i, namely can be considered that this independent variable has a certain impact to target variable tool, if the absolute value of the coefficient of maximum absolute value and the second largest coefficient of absolute value differs by more than equal 1/i, then clearly can determine that the maximum independent variable of absolute coefficient is the main affecting factors of target variable, otherwise just the independent variable of these two larger coefficients is defined as main affecting factors.

Beneficial effect

The invention provides a kind of method of determining to dry the silk effects of process parameters factor based on PLS, by the target variable in selected baking silk technological process and independent variable, gather the historical data of target variable and independent variable; Adopt Euclidean distance method to screen historical data, reject the abnormal data in historical data, obtain sample data; Calculate the correlation coefficient matrix between independent variable x and independent variable x, target variable y and target variable y and independent variable x and target variable y drying silk technique; Observe the correlativity size between independent variable and target variable, adopt least square method PLS to set up the PLS model of independent variable x and target variable y, calculate the size of the factor of influence of the independent variable x obtaining each target variable y of impact according to model.The method is simple, be easy to realize, and application this method substitutes artificial experience and judges, obtains the size of factor of influence fast and accurately.

Accompanying drawing explanation

Fig. 1 is schematic flow sheet of the present invention;

Fig. 2 is the histogram of each independent variable to the factor of influence size of target variable y1;

Fig. 3 is the histogram of each independent variable to the factor of influence size of target variable y2.

Embodiment

Below in conjunction with accompanying drawing, the present invention will be further described.

Produce certain model cigarette for example with certain cigarette factory, a kind of method determining the baking silk effects of process parameters factor based on PLS, its schematic flow sheet as shown in Figure 1.

1) target variable y and independent variable x determines and the collection of historical data

Drying technological parameter involved in silk technique a lot, comprise the process data (HT vapor pressure, HT steam flow, inlet water grade) of HT section; Drum inlet process data (drum inlet hot blast temperature, drum inlet hot blast air quantity etc.); Barrel supplemental characteristic (barrel temperature, barrel vapor pressure etc.); Humidity discharging wind supplemental characteristic (humidity discharging throttle opening, discharge cover negative pressure etc.) and pipe tobacco discharging supplemental characteristic (discharging moisture content of cut tobaccos, discharging pipe tobacco temperature etc.).The present embodiment is selected with the water percentage of discharging pipe tobacco and outlet temperature as target variable (y ₁, y ₂), since HT entrance moisture, HT vapor pressure, HT steam flow, hot blast temperature, hot blast air door, discharge cover pressure, cylinder wall temperature, storage heater steam flow, pipe tobacco instantaneous delivery be independent variable (x ₁, x ₂... x ₉);

From historical data base, extract the sample data of 2 months (2012-3-01 ~ 2012-4-25), the sample frequency of data is set to 60s.These data comprise target variable discharging moisture content of cut tobaccos and outlet temperature, form original object variable sample Y; Extract whole argument data, these data form original argument's sample X simultaneously;

2) screening sample

Owing to may there is the abnormal occurrence such as equipment failure, shutdown in history production run, may there is the sample of improper production in the sample data of extraction, and the existence of these samples can affect the final analysis result of PLS, therefore must screen sample.This programme adopts Euclidean distance method (Euclideandistance) to delete outliers, and the data after screening sample Variable Selection have n group, independent variable sample X _nj(i=1,2 ..., n; J=1,2 ..., 9), target variable sample Y _nj(i=1,2 ..., n; J=1,2).

3) correlation coefficient matrix calculates

Calculate the correlation coefficient matrix between independent variable x and independent variable x, target variable y and target variable y and independent variable x and target variable y drying silk technique.

Related coefficient is also known as linearly dependent coefficient, and it is the index weighing linear correlation degree between variable, and sample correlation coefficient r represents, Calculation of correlation factor formula is as follows:

$r=\frac{\mathrm{\Σ}(X-\stackrel{\‾}{X})(Y-\stackrel{\‾}{Y})}{\sqrt{\mathrm{\Σ}{(X-\stackrel{\‾}{X})}^2\mathrm{\Σ}{(Y-\stackrel{\‾}{Y})}^2}}=\frac{\mathrm{\ΣXY}-\frac{\mathrm{\ΣX}\·\mathrm{\ΣY}}{n}}{\sqrt{[\mathrm{\Σ}{X}^2-\frac{{\left(\mathrm{\ΣX}\right)}^2}{n}][\mathrm{\Σ}{Y}^2-\frac{{\left(\mathrm{\ΣY}\right)}^2}{n}]}}$

Table 1: dry silk process parameter table

Between each variable that the sample data after screening based on step 2 calculates, related coefficient is as shown in table 2, table 2 represents the correlation analysis matrix dried between silk technological process important parameter, numerical value in form represents the correlativity of two parameters in the gauge outfit of corresponding row and column, negative number representation negative correlation (namely another parameter value of a parameter value increase correspondence can reduce), in table, gray cell lattice institute column data absolute value is greater than 50, therefrom can find out that between corresponding two parameters, correlativity is stronger.

Table 2: dry silk technological parameter correlation matrix (%)

4) PLS modeling is carried out

Utilize partial least square method to carry out computing to data, by cross validation test method, determine that PLS component number is 5.Iterative algorithm in partial least square method, obtains the load factor matrix Q of objective matrix Y ^twith the loading matrix P of technological parameter matrix X ^t:

$W=\left[\begin{array}{ccccccccc}0.47& -0.17& -0.77& -0.22& -0.03& 0.27& -0.06& -0.11& 0.15\\ -0.12& 0.14& -0.01& 0.03& -0.54& -0.16& -0.57& -0.56& 0.05\\ 0.13& -0.24& -0.2& 0.31& -0.06& -0.59& -0.38& 0.54& 0.05\\ -0.05& 0.7& -0.03& -0.18& -0.12& 0.22& -0.2& 0.47& 0.39\\ -0.1& -0.11& 0.01& -0.03& 0& 0.5& -0.47& 0.27& -0.66\end{array}\right]$

$Q^T=\left[\begin{array}{cc}0.47& -0.88\\ -0.04& -1\\ 0.92& -0.4\\ 0.62& 0.79\\ 0.98& -0.19\end{array}\right]$

$P^T=\left[\begin{array}{ccccccccc}63.05& -28.29& -87.8& -32.29& 34.62& 51.18& 36.97& 13.82& 12.21\\ -23.73& 58.02& 1.67& -9.28& -91.71& -4.22& -88.03& -86.37& 25.69\\ 16.07& -25.3& -17.68& 30.34& -4.97& -65.25& -16.64& 33.15& 10.25\\ -8.9& 54.64& -0.39& -18.41& -5.66& 13.42& -14.43& 35.98& 56.33\\ -9.99& -7.59& -3.33& -3.73& -16.24& 40.41& -28.7& 27.53& -64.56\end{array}\right]$

Technological parameter matrix X obtains resolute t ₁-t ₅just can calculate according to the following formula of formula:

t ₁=63.05x ₁-28.29x ₂-87.8x ₃-32.29x ₄+34.62x ₅+51.18x ₆+36.97x ₇+13.82x ₈+12.21x ₉

t ₂=-23.73x ₁+58.02x ₂+1.67x ₃-9.28x ₄-91.71x ₅-4.22x ₆-88.03x ₇-86.37x ₈+25.69x ₉

t ₃=16.07x ₁-25.3x ₂-17.68x ₃+30.34x ₄-4.97x ₅-65.25x ₆-16.64x ₇+33.15x ₈+10.25x ₉

t ₄=-8.9x ₁+54.64x ₂-0.39x ₃-18.41x ₄-5.66x ₅+13.42x ₆-14.43x ₇+35.98x ₈+56.33x ₉

t ₅=-9.99x ₁-7.59x ₂-3.33x ₃-3.73x ₄-16.24x ₅+40.41x ₆-28.7x ₇+27.53x ₈-64.56x ₉

Equally, objective matrix Y resolute u ₁-u ₅for:

u ₁=0.47y ₁-0.88y ₂

u ₂=-0.04y ₁-y ₂

u ₃=0.92y ₁-0.4y2

u ₄=0.62y ₁+0.79y ₂

u ₅=0.98y1-0.19y ₂

Two inside correlation matrixes obtained between resolute are:

$V=\left[\begin{array}{ccccc}70.61& 0& 0& 0& 0\\ 0& 33.63& 0& 0& 0\\ 0& 0& 27.18& 0& 0\\ 0& 0& 0& 17.39& 0\\ 0& 0& 0& 0& 8.33\end{array}\right]$

Thus two internal relations formulas obtained between resolute are expressed as:

u ₁=70.61t ₁+eu ₂=33.63t ₂+eu ₃=27.18t ₃+eu ₄=17.39t ₄+eu ₅=8.33t5+e

E is regression residuals.

Finally can calculate the relation B of target vector and technological parameter vector:

$B=\left[\begin{array}{ccccccccc}0.18& -0.04& -0.33& 0& -0.09& -0.04& -0.19& 0.24& 0.05\\ -0.33& 0.16& 0.54& 0.07& 0.17& -0.04& 0.17& 0.23& -0.02\end{array}\right]$

y ₁=0.18x ₁-0.04x ₂-0.33x ₃+0x ₄-0.09x ₅-0.04x ₆-0.19x ₇+0.24x ₈+0.05x ₉

y2=-0.33x ₁+0.16x ₂+0.54x ₃+0.07x ₄+0.17x ₅-0.04x ₆+0.17x ₇+0.23x ₈-0.02x ₉

The result of calculation obtained after modeling as shown in Figures 2 and 3, Fig. 2 and Fig. 3 presents the moisture content of outlet that obtains after PLS analysis modeling and the related coefficient histogram between outlet temperature and factor of influence respectively, can recognize to quicklook that from figure the factor of influence that those related coefficient absolute values are greater than 0.2 is the principal element affecting target variable, also can see the influence degree of each influence factor to target variable simultaneously.

Utilize PLS to this section of historical data modeling, the principal element that obtaining affects moisture content of outlet mass property is followed successively by: HT section steam flow, hot blast air quantity, barrel temperature, HT section entrance moisture, hot blast temperature; The principal element affecting outlet temperature mass property is followed successively by: HT section steam flow, HT section entrance moisture, hot blast air quantity, hot blast temperature, barrel temperature, HT section vapor pressure, discharge cover pressure.

After PLS analyzes, obtain analysis result, then contrast with the process experiences of technologist's reality.The conclusion that PLS analysis modeling obtains is substantially identical with the actual condition of production.In sum, this patent can confirm to affect the mass property factor of drying silk, reliable results fast and accurately.

Claims (1)

1. determine a method of drying the silk technogenic influence factor based on PLS, it is characterized in that, comprise the following steps:

Step 1: independent variable x and the target variable y of silk technique are dried in setting, independent variable x chooses from HT vapor pressure, HT steam flow, entrance moisture, drum inlet hot blast temperature, drum inlet hot blast air quantity, barrel temperature, barrel vapor pressure, humidity discharging throttle opening and discharge cover negative pressure, target variable y dries silk water percentage from discharging, discharging is dried silk temperature and humidity discharging humidity and chosen, gather the historical data of independent variable x and target variable y, wherein the sample frequency of historical data is T;

Step 2: adopt Euclidean distance method to screen historical data, reject the abnormal data in historical data, obtain n group sample data, sample data forms target variable matrix Y _njwith independent variable matrix X _ni, target variable Y has j, and independent variable X has i;

Described abnormal data refers to the baking silk process data drying and gather when silk process equipment breaks down;

Step 3: the sample data after the screening obtained according to step 2, sets up independent variable X _{n × i}with target variable Y _{n × j}partial least square method PLS model, calculate the size of factor of influence of each independent variable x of each target variable y of impact according to PLS model, its specific operation process is as follows:

1) the PLS component number in partial least square method PLS and cross validation test method determination sample data is utilized to be h;

2) adopt the iterative algorithm in partial least square method, be obtained from matrix of variables X _niweight matrix W, objective matrix Y _njload factor matrix Q ^twith independent variable matrix X _niloading matrix P ^t:

$W=\left[\begin{array}{ccc}{w}_{11}& ...& w_{1i}\\ w_{21}& ...& w_{2i}\\ ...& ...& ...\\ ...& ...& ...\\ w_{h1}& ...& w_{hi}\end{array}\right]$

$O^T=\left[\begin{array}{ccc}{a}_{11}& ...& a_{1j}\\ a_{21}& ...& a_{2j}\\ ...& ...& ...\\ ...& ...& ...\\ a_{h1}& ...& a_{hj}\end{array}\right]$ $P^T=\left[\begin{array}{ccc}{c}_{11}& ...& c_{1i}\\ c_{21}& ...& c_{2i}\\ ...& ...& ...\\ ...& ...& ...\\ c_{h1}& ...& c_{hi}\end{array}\right]$

3) that determines independent variable matrix X obtains resolute t _hwith objective matrix Y resolute u _h:

Independent variable matrix X _{n × i}resolute t _hfor:

t _h＝c _h1x ₁+c _h2x ₂+…+c _hix _i

Objective matrix Y _{n × j}resolute u _hfor:

u _h＝a _h1y ₁+…+a _hjy _j

Wherein, h is the component number of sample data, and i is independent variable number, and j is target variable number;

4) to independent variable matrix X _{n × i}weight matrix W be normalized calculating, obtain two inside correlation matrixes between resolute are V:

$V=\left[\begin{array}{cccc}{v}_1& 0& ...& 0\\ 0& ...& ...& ...\\ ...& ...& ...& ...\\ ...& ...& ...& 0\\ 0& ...& 0& v_h\end{array}\right]$

Thus two internal relations formulas obtained between resolute are expressed as:

U ₁=v ₁t ₁+ e, u ₂=v ₂t ₂+ e, u ₃=v ₃t ₃+ e ... u _h=v _ht _h+ e, e are regression residuals;

5) according to B=V (P ^tv) ^-1q ^tcalculate the matrix of coefficients B in Y=XB:

$B=\left[\begin{array}{ccc}{b}_{11}& ...& b_{1i}\\ b_{21}& ...& b_{2i}\\ ...& ...& ...\\ ...& ...& ...\\ b_{j1}& ...& b_{ji}\end{array}\right]$

Thus the relation finally obtained between independent variable x and target variable y is as follows:

y ₁＝b ₁₁x ₁+b ₁₂x ₂+b ₁₃x ₃+……+b _1ix _i

y ₂＝b ₂₁x ₁+b ₂₂x ₂+b ₂₃x ₃+……+b _2ix _i

……

y _j＝b _j1x ₁+b _j2x ₂+b _j3x ₃+……+b _jix _i

Step 4: extract the factor of influence of independent variable as corresponding target variable that independent variable absolute coefficient in above-mentioned formula is greater than 1/i, i is the number of independent variable X.