CN104298706A - Transmission tower material actual strength computing method based on data mining - Google Patents

Abstract

The invention discloses a transmission tower material actual strength computing method based on data mining. A rough set theory is used for reducing a pole and tower strength influence factor set; a data mining method is used for mining useful data from a large number of historical data to establish a learning set; and according to the learning set, a gradient descent method is used for constant iterative learning to determine factor set weight, and through the obtained weight value, the degeneration rate of the actual strength of tower materials is subjected to comprehensive evaluation and computing. The problems of typical uncertainty and nonlinearity that during an evolutionary process, the validity affecting tower material strength factors is not clear can be solved through the method. The method can provide an important scientific criterion for transmission tower structure safety evaluation.

Description

A kind of transmission tower tower material actual strength computing method based on data mining

Technical field

The present invention relates to electric power pylon safety appraisement of structure field, be specially adapted to the transmission tower safety appraisement of structure of long-time running under Complex Natural Environment.

Background technology

The essence of line construction safety is a uncertainty, nonlinear state space evolutionary process, shape State evolution (transfer) process has randomness, the characteristic information wherein characterizing shaft tower tower material actual machine intensity has inexactness, the effect validity of influence factor is also unintelligible, definition and the extension of running status have ambiguity, the expertise of state-evaluation has incompleteness, so be a complicated uncertain problem to the evaluation of circuit tower material intensity with calculating.

Data mining can disclose implicit, previously unknown and have the non-trivial process of the information of potential value from the mass data of database.Data mining is a kind of decision support processes, it is mainly based on artificial intelligence, machine learning, pattern-recognition, statistics, database, visualization technique etc., analyze data increasingly automatedly, make the reasoning of inductive, therefrom excavate potential pattern, aid decision making person's adjustable strategies, reduces risks, and makes correct decision-making.

Because the factor affecting tower material intensity is a lot, and traditional evaluation method needs to calculate whole evaluation index, so calculate more complicated, and the serious distortion phenomenon that may cause because each factor weight is little or multi-peak phenomenon.The present invention proposes one and first carries out yojan based on rough set theory to index set, adopts the method for data mining to carry out the method for tower material actual strength calculating after yojan.Institute's extracting method effectively can solve the transmission tower safety appraisement of structure problem of long-time running under Complex Natural Environment.

Summary of the invention

The object of the invention is to: propose a kind of transmission tower tower material actual strength computing method based on data mining, can be assessment electric power line tower structure and necessary criterion is provided safely.

Step 1: index set yojan;

Step 2: the structure of data mining study collection;

Step 3: set of factors weight is determined;

Step 4: actual strength is assessed.

Index set yojan in described step 1, mainly to the index set by meteorologic district condition, sub-strength damage, the stress of conductor and mechanical vibration three major types factors composition, three class influence factor set representations are U={U ₁, U ₂, U ₃.Wherein: U ₁={ u ₁₁, u ₁₂, u ₁₃, u ₁₄, u ₁₅, u ₁₁for wind speed (most strong wind), u ₁₂for atmospheric temperature (lowest temperature), u ₁₃for average temperature of the whole year, u ₁₄for ice covering thickness (the thickest icing), u ₁₅for number of days in thunderstorm day in year.U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, u ₂₆, u ₂₁for working time, u ₂₂for bending reparation number of times, u ₂₃for slight crack repairs number of times, u ₂₄for thunder and lightning or fault current damage number of times, u ₂₅to attach most importance to icing times of fatigue, u ₂₆for on average running stress/maximum operation stress.U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄, u ₃₁for wire division number, u ₃₂for wind direction and circuit angle, u ₃₃for ground surface degree of roughness, u ₃₄for steel corrosion amount.Use rough set to carry out index set yojan, Attribute Significance is defined as:

U/R＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₁})＝{{1,3,5,7,8},{2,4}}

U/(R-{u ₁₂})＝{{1,2,4,5,7},{3,6,8}}

U/(R-{u ₁₃})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₄})＝{{1,5,7},{2,4},{3,6,8}}

U/(R-{u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₃,u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/R≠U/(R-{u ₁₁})

U/R≠U/(R-{u ₁₂})

U/R≠U/(R-{u ₁₄})

U/R＝U/(R-{u ₁₃})＝U/(R-{u ₁₅})＝U/(R-{u ₁₃,u ₁₅})

Wherein R is the set of the property value of factor.Known index u is calculated through Attribute Significance yojan ₁₃, u ₁₅be redundancy, in like manner, carry out Attribute Significance yojan to sub-strength damage factor and conductor galloping and aeolian vibration factor respectively, obtaining final evaluation index is: U={U ₁, U ₂, U ₃, wherein U ₁={ u ₁₁, u ₁₂, u ₁₄, U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄.

The structure of the data mining study collection in described step 2, adopt K-VNN search strategy, in system existing N group data, find the most similar data (k<<N) of k group, the actual strength degradation ratio of the shaft tower tower material that the present invention proposes and influence factor collection U present nonlinear relationship:

η＝f(U)+ε(t)

Wherein, η is degradation ratio, the white noise that ε (t) is zero-mean, and f () is unknown nonlinear function.For the data that there is N group influence factor and degradation ratio in current t, there is influence factor information U (t), adopt K-VNN search strategy, in system existing N group data, find the most similar data (k<<N) of k group, specific as follows:

1, as cos β (U (i), U (t)) <0, then think that this U (i) deviates from current input U (t), be unfavorable for modeling, abandon this data;

2 otherwise, with the index core of U (i) and U (t) and included angle cosine weighting sum selection criterion, namely

D(U(i),U(t))＝α·e ^{-d(U(i),U(t))}+(1-α)·cosβ(U(i),U(t))

In formula:

$\left\{\begin{array}{c}\cos \mathrm{\&beta;}(U\left(i\right),U\left(t\right))=\frac{U^T\left(i\right)U\left(t\right)}{{\left|\right|U\left(i\right)\left|\right|}_2\&CenterDot;{\left|\right|U\left(t\right)\left|\right|}_2}\\ d(U\left(i\right),U\left(t\right))={\left|\right|U\left(i\right)-U\left(i\right)\left|\right|}_2,\mathrm{\&alpha;}\&Element;\left[\mathrm{0,1}\right]\end{array}\right.$

α is weighting factor.Weighting selection criterion D (U (i), U (t)) directly reflects the similarity of U (i) and U (t).If two information vectors the closer to, then d is less, and cos β is also larger, thus D (U (i), U (t)) is also larger.Like this, in existing data message, the k group data selecting D () value maximum, by descending sort, learning of structure collection:

{(U(1),η(1)),…,(U(k),η(k))}

D(U(1),U(t))>…>D(U(k),U(t)).

Set of factors weight in described step 3 is determined to obtain the influence degree of certain factor to shaft tower tower material actual strength degradation ratio, by probabilistic factor sharpening, actual strength degradation ratio and the influence factor collection U local linear relation of the shaft tower tower material of the present invention's proposition are as follows:

η＝f _θ(U)＝U ^Tθ

Wherein: θ represents set of factors weighted value, and T represents transposition.Because current and data are in different working points, the packing density meeting present operating point U (t) may be different, data amount check for modeling is also indefinite, that is: modeling neighborhood value size is variable, in order to obtain best set of factors weight vectors θ, reduce calculated amount simultaneously, the variation range k ∈ [k of neighborhood can be preset _m, k _m] (k _m<k _m), calculating the set of factors weight vectors θ of neighbour k+1 _k+1time, directly utilize the set of factors weight vectors θ of neighbour k _k, first provide an error functions,

$\underset{\mathrm{\&theta;}}{\min }J\left(\mathrm{\&theta;}\right)=\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2$

Adopt gradient descent method, calculate set of factors weight vectors as follows,

$\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)=\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2=(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

${\mathrm{\&theta;}}^{k+1}={\mathrm{\&theta;}}^k-\mathrm{\&alpha;}\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)={\mathrm{\&theta;}}_i-\mathrm{\&alpha;}(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

Obtaining is the model θ of neighbour k+1 _k+1, meanwhile, what also can obtain neighbour k+1 goes a transposition error value:

$e_{k+1,j}^{\mathrm{loo}}=\mathrm{\&eta;}\left(j\right)-\mathrm{\&alpha;}{\mathrm{\&theta;}}_{k+1}^{-j},j=\mathrm{1,2}...,k+1.$

In formula: represent in k+1 group data, with removing a jth model that data gained arrives; represent actual column material strength degradation rate η (j) and model error between the predicted value obtained.

Like this, what can obtain neighbour k+1 removes a transposition error collection all side and these ?

$E^{\mathrm{loo}}(k+1)=\frac{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j{\left(e_{k+1,j}^{\mathrm{loo}}\right)}^2}{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j}.$

In formula: weighting factor the each U (j) of direct reflection go a transposition error to E ^loo(k+1) " contribution " size.The closer to the U (j) of U (t), its " contribution " is larger, otherwise less.Now, if

E ^loo(k+1)>E ^loo(k),k+1∈[k _m,k _M].

Then think and stop model " variation " returning calculating, and with model θ _kas the best model of system current time.Otherwise, by the model adopting gradient descent method to obtain, concentrate from study and choose the information vector made new advances, continue iteration, until k=k _mtill.Like this, the quality of partial model can be judged in time, obtain the best Local Linear Model meeting current time influence factor and degradation ratio relation.

Safety assessment in described step 4, according to the influence factor of foundation and the best Local Linear Model of degradation ratio relation, calculates the degradation ratio of tower material actual strength, and the degradation ratio of the calculating shaft tower tower material actual strength that the present invention proposes is as follows,

η(t)＝＝U ^T(t)θ _k。

Technique effect of the present invention:

The application is compared to classic method, not only make index be simplified, and make the weight allocation of each factor utilize a large amount of history image data, the discrimination of its evaluation result is obvious, result gears to actual circumstances more, the reflected appraisal result making evaluation index more effective.

Accompanying drawing explanation

Fig. 1 is the schematic diagram of the actual strength computing method of long-time running power transmission line column material.

Fig. 2 is that degradation ratio calculates comparative evaluation result figure.

Embodiment

As shown in Figure 1, the method mainly comprises the steps: the mentality of designing of a kind of transmission tower tower material actual strength computing method based on data mining of the present invention

Step 1: index set yojan;

Step 2: the structure of data mining study collection;

Step 3: set of factors weight is determined;

Step 4: actual strength is assessed.

The implementation process of each step is as follows:

Electric power line pole tower tower material long-term field runs, and bear nature invasion and attack, electric discharge and artificial outside destroy, wire actual machine intensity relative theory value can decline.Consider intensity effect factor and pass judgment on knowledge to present nonlinear relationship, based on data digging method, shaft tower tower material actual strength is evaluated and calculated.

The K-VNN method of one of data mining is based on " similar input produces similar output " principle, implementation is generally by sample data memory in memory, then according to input point, in sample data, find data similarly, obtain the corresponding output of this input according to these sample datas.Therefore, it is also referred to as " based on mnemonic learning ".The criterion describing input data and the sample data degree of association generally adopts distance function, and those namely nearest with input point data have the degree of association higher with it.The basic process of intensity evaluation as shown in Figure 1.The present invention establishes local strength's evaluation model, comprises the foundation of set of factors, weight study collection, evaluation model.Hereafter for strain resistant shaft tower tower material, provide its actual strength calculation method.

1) tower material intensity evaluation set of factors index system is set up;

By in power supply enterprise's investigation, according to the evaluation index affecting the various factors of tower material intensity that person skilled provides, in conjunction with a large amount of real data, draw 110kV electric pressure line steel tower intensity evaluation basic index system.Basic index system is primarily of meteorologic district condition, sub-strength damage, the stress of conductor and mechanical vibration three major types factors composition.

Set of factors is: U={U ₁, U ₂, U ₃, wherein U ₁={ u ₁₁, u ₁₂, u ₁₃, u ₁₄, u ₁₅, U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, u ₂₆, U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄.

1. meteorologic district condition U ₁:

Wind speed (most strong wind) u ₁₁, atmospheric temperature (lowest temperature) u ₁₂, average temperature of the whole year u ₁₃, ice covering thickness (the thickest icing) u ₁₄, thunderstorm day in year number of days u ₁₅.

2. sub-strength damage U ₂:

Working time u ₂₁, bending reparation number of times u ₂₂, slight crack repairs number of times u ₂₃, thunder and lightning or fault current damage number of times u ₂₄, re-cover ice times of fatigue u ₂₅, on average run stress/maximum operation stress u ₂₆.

3. the stress of conductor and mechanical vibration U ₃:

Wire division number u ₃₁, wind direction and circuit angle u ₃₂, ground surface degree of roughness u ₃₃, steel corrosion amount u ₃₄.

Because the value of each factor is different, and comparatively complicated, in order to the convenience that further evaluation calculates, at this, data normalization is carried out to basic index.Five grades are divided into according to the influence degree difference of factor of evaluation to tower material intensity, be respectively I, II, III, IV, V, grade I represents very little on the impact of tower material actual strength, grade II represents less on the impact of tower material actual strength, grade III represents medium on the impact of tower material actual strength, grade IV represents comparatively large on the impact of tower material actual strength, and grade V represents very large on the impact of tower material actual strength.Set up meteorologic district for meteorologic district below and evaluate quantitative criteria.

Set up the evaluation criterion of meteorologic district condition by national typical meteorological condition area storehouse, show that quantitative criteria is evaluated in meteorologic district, in table 1.

Quantitative criteria is evaluated in table 1 meteorologic district

2) rough set carries out attribute reduction

The many factors of basic index system, may the problem of existing factor redundancy, on the basis of not effect appraise result, in order to not allow evaluation procedure complicated, will carry out yojan to basic index system.Test sample data are as shown in table 2, and for the ease of computational short cut, the desired value using II is as the threshold values of each initial evaluation index, the desired value meeting II is then 1, otherwise is 0, then carry out Data Discretization by table 2 and table 3 data, just initial evaluation indication information can be drawn, in table 3.

Table 2 test sample data

Table 3 sample data discretize

According to rough set theory, his-and-hers watches 4 carry out Attribute Significance yojan:

U/R＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₁})＝{{1,3,5,7,8},{2,4}}

U/(R-{u ₁₂})＝{{1,2,4,5,7},{3,6,8}}

U/(R-{u ₁₃})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₄})＝{{1,5,7},{2,4},{3,6,8}}

U/(R-{u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₃,u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/R≠U/(R-{u ₁₁})

U/R≠U/(R-{u ₁₂})

U/R≠U/(R-{u ₁₄})

U/R＝U/(R-{u ₁₃})＝U/(R-{u ₁₅})＝U/(R-{u ₁₃,u ₁₅})

Through Attribute Significance yojan, calculate known index u ₁₃, u ₁₅it is redundancy.In like manner, carry out Attribute Significance yojan to sub-strength damage factor and the stress of conductor and mechanical vibration factor respectively, obtaining final evaluation index is: U={U ₁, U ₂, U ₃, wherein U ₁={ u ₁₁, u ₁₂, u ₁₄, U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄.

3) structure of study collection

Consider that unknown entering singly goes out Nonlinear Mapping f:R more ⁿ→ R, assuming that the observable inputoutput data of system can be obtained: and this group data existence function relation:

y _i＝h(x _i)+ε

In formula: X ∈ R ⁿit is independent variable; y _i∈ R is dependent variable: ε _i∈ R is zero-mean and variance is σ ²independent random distribution variable.Problem is any vectorial X for the input space _q, according to the existing data set of system, a mapping can be set up, and by this mapping, the system that obtains is estimated to export accordingly this problem can be summed up as the optimization problem solved below

$\min \underset{(X_i,Y_i)\&Element;{\mathrm{\&Omega;}}_k}{\mathrm{\&Sigma;}}{(y_i-h(X_i,\mathrm{\&theta;}))}^2{w}_i$

In formula: Ω _kfor distance X _kthe local space that k nearest sample is formed; H () is the nonlinear mapping function describing input and output vector; w _ifor weights, the sample data in expression local space is to the influence degree of output vector, and the impact (or contribution degree) that sample datas different in local space exports system is different.Intuitively, the output vector value corresponding to vectorial those the nearest samples input of distance input can reflect the output of current input quantity, and this is in fact also the cardinal rule of Lazy learning method: similar input produces similar output.

Use above algorithm principle, adopt K-VNN search strategy, in system existing N group data, find the most similar data (k<<N) of k group, the actual strength degradation ratio of the shaft tower tower material that the present invention proposes and influence factor collection U present nonlinear relationship:

η＝f(U)+ε(t)

Wherein, η is degradation ratio, the white noise that ε (t) is zero-mean, and f () is unknown nonlinear function.For the data that there is N group influence factor and degradation ratio in current t, there is influence factor information U (t), adopt K-VNN search strategy, in system existing N group data, find the most similar data (k<<N) of k group, specific as follows:

2, as cos β (U (i), U (t)) <0, then think that this U (i) deviates from current input U (t), be unfavorable for modeling, abandon this data;

2 otherwise, with the index core of U (i) and U (t) and included angle cosine weighting sum selection criterion, namely

D(U(i),U(t))＝α·e ^{-d(U(i),U(t))}+(1-α)·cosβ(U(i),U(t))

In formula:

$\left\{\begin{array}{c}\cos \mathrm{\&beta;}(U\left(i\right),U\left(t\right))=\frac{U^T\left(i\right)U\left(t\right)}{{\left|\right|U\left(i\right)\left|\right|}_2\&CenterDot;{\left|\right|U\left(t\right)\left|\right|}_2}\\ d(U\left(i\right),U\left(t\right))={\left|\right|U\left(i\right)-U\left(i\right)\left|\right|}_2,\mathrm{\&alpha;}\&Element;\left[\mathrm{0,1}\right]\end{array}\right.$

α is weighting factor.Weighting selection criterion D (U (i), U (t)) directly reflects the similarity of U (i) and U (t).If two information vectors the closer to, then d is less, and cos β is also larger, thus D (U (i), U (t)) is also larger.Like this, in existing data message, the k group data selecting D () value maximum, by descending sort, learning of structure collection:

{(U(1),η(1)),…,(U(k),η(k))}

D(U(1),U(t))>…>D(U(k),U(t)).

4) set of factors weight is determined

Actual strength degradation ratio and the influence factor collection U local linear relation of shaft tower tower material are as follows:

η＝f _θ(U)＝U ^Tθ

Wherein: θ represents set of factors weighted value, and T represents transposition.Because current and data are in different working points, the packing density meeting present operating point U (t) may be different, data amount check for modeling is also indefinite, that is: modeling neighborhood value size is variable, in order to obtain best set of factors weight vectors θ, reduce calculated amount simultaneously, the variation range k ∈ [k of neighborhood can be preset _m, k _m] (k _m<k _m), calculating the set of factors weight vectors θ of neighbour k+1 _k+1time, directly utilize the set of factors weight vectors θ of neighbour k _k, first provide an error functions,

$\underset{\mathrm{\&theta;}}{\min }J\left(\mathrm{\&theta;}\right)=\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2$

Adopt gradient descent method, calculate set of factors weight vectors as follows,

$\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)=\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2=(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

${\mathrm{\&theta;}}^{k+1}={\mathrm{\&theta;}}^k-\mathrm{\&alpha;}\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)={\mathrm{\&theta;}}_i-\mathrm{\&alpha;}(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

Obtaining is the model θ of neighbour k+1 _k+1, meanwhile, what also can obtain neighbour k+1 goes a transposition error value:

$e_{k+1,j}^{\mathrm{loo}}=\mathrm{\&eta;}\left(j\right)-\mathrm{\&alpha;}{\mathrm{\&theta;}}_{k+1}^{-j},j=\mathrm{1,2}...,k+1.$

In formula: represent in k+1 group data, with removing a jth model that data gained arrives; represent actual column material strength degradation rate η (j) and model error between the predicted value obtained.

Like this, what can obtain neighbour k+1 removes a transposition error collection all side and these ?

$E^{\mathrm{loo}}(k+1)=\frac{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j{\left(e_{k+1,j}^{\mathrm{loo}}\right)}^2}{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j}.$

In formula: weighting factor the each U (j) of direct reflection go a transposition error to E ^loo(k+1) " contribution " size.The closer to the U (j) of U (t), its " contribution " is larger, otherwise less.Now, if

E ^loo(k+1)>E ^loo(k),k+1∈[k _m,k _M].

Then think and stop model " variation " returning calculating, and with model θ _kas the best model of system current time.Otherwise, by the model adopting gradient descent method to obtain, concentrate from study and choose the information vector made new advances, continue iteration, until k=k _mtill.Like this, the quality of partial model can be judged in time, obtain the best Local Linear Model meeting current time influence factor and degradation ratio relation.So this Local Linear Model may be used for the degradation ratio calculating shaft tower tower material actual strength, i.e. η (t)==U ^t(t) θ _k.

For the strain resistant steel tower of certain Utilities Electric Co., transfer the detailed weather data of nearly 5 years, and the detailed service data that steel tower puts into operation, through native system again inverse modeling, system security level is Failure risk state.In addition, can find out that the resultant error that the steel tower strength degradation rate that calculated by institute of the present invention extracting method and manual analysis obtain is less by emulation, may be used for replacing artificial calculating, its result of calculation as shown in Figure 2.

The content be not described in detail in this manual belongs to the known technology of those skilled in the art.

Claims (5)

1., based on transmission tower tower material actual strength computing method for data mining, it is characterized in that: comprise the steps:

Step 1: index set yojan;

Step 2: the structure of data mining study collection;

Step 3: set of factors weight is determined;

Step 4: actual strength is assessed.

2. a kind of transmission tower tower material actual strength computing method based on data mining according to claim 1, it is characterized in that: the index set yojan in described step 1, main to the index set by meteorologic district condition, sub-strength damage, the stress of conductor and mechanical vibration three major types factors composition, three class influence factor set representations are U={U ₁, U ₂, U ₃.Wherein: U ₁={ u ₁₁, u ₁₂, u ₁₃, u ₁₄, u ₁₅, u ₁₁for wind speed (most strong wind), u ₁₂for atmospheric temperature (lowest temperature), u ₁₃for average temperature of the whole year, u ₁₄for ice covering thickness (the thickest icing), u ₁₅for number of days in thunderstorm day in year.U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, u ₂₆, u ₂₁for working time, u ₂₂for bending reparation number of times, u ₂₃for slight crack repairs number of times, u ₂₄for thunder and lightning or fault current damage number of times, u ₂₅to attach most importance to icing times of fatigue, u ₂₆for on average running stress/maximum operation stress.U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄, u ₃₁for wire division number, u ₃₂for wind direction and circuit angle, u ₃₃for ground surface degree of roughness, u ₃₄for steel corrosion amount.Use rough set to carry out index set yojan, Attribute Significance is defined as:

U/R＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₁})＝{{1,3,5,7,8},{2,4}}

U/(R-{u ₁₂})＝{{1,2,4,5,7},{3,6,8}}

U/(R-{u ₁₃})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₄})＝{{1,5,7},{2,4},{3,6,8}}

U/(R-{u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/(R-{u ₁₃,u ₁₅})＝{{1,7},{2,4},{3,6,8},{5}}

U/R≠U/(R-{u ₁₁})

U/R≠U/(R-{u ₁₂})

U/R≠U/(R-{u ₁₄})

U/R＝U/(R-{u ₁₃})＝U/(R-{u ₁₅})＝U/(R-{u ₁₃,u ₁₅})

Wherein R is the set of the property value of factor.Known index u is calculated through Attribute Significance yojan ₁₃, u ₁₅be redundancy, in like manner, carry out Attribute Significance yojan to sub-strength damage factor and conductor galloping and aeolian vibration factor respectively, obtaining final evaluation index is: U={U ₁, U ₂, U ₃, wherein U ₁={ u ₁₁, u ₁₂, u ₁₄, U ₂={ u ₂₁, u ₂₂, u ₂₃, u ₂₄, u ₂₅, U ₃={ u ₃₁, u ₃₂, u ₃₃, u ₃₄.

3. a kind of transmission tower tower material actual strength computing method based on data mining according to claim 1, it is characterized in that: the structure of the data mining study collection in described step 2, adopt K-VNN search strategy, in system existing N group data, find the most similar data k<<N of k group, actual strength degradation ratio and the influence factor collection U of shaft tower tower material present nonlinear relationship:

η＝f(U)+ε(t)

Wherein, η is degradation ratio, the white noise that ε (t) is zero-mean, and f () is unknown nonlinear function.For the data that there is N group influence factor and degradation ratio in current t, there is influence factor information U (t), adopt K-VNN search strategy, in system existing N group data, find the most similar data (k<<N) of k group, specific as follows:

As cos β (U (i), U (t)) <0, then think that this U (i) deviates from current input U (t), be unfavorable for modeling, abandon this data;

Otherwise, with the index core of U (i) and U (t) and included angle cosine weighting sum selection criterion, namely

D(U(i),U(t))＝α·e ^{-d(U(i),U(t))}+(1-α)·cosβ(U(i),U(t))

In formula:

$\left\{\begin{array}{c}\cos \mathrm{\&beta;}(U\left(i\right),U\left(t\right))=\frac{U^T\left(i\right)U\left(t\right)}{{\left|\right|U\left(i\right)\left|\right|}_2\&CenterDot;{\left|\right|U\left(t\right)\left|\right|}_2}\\ d(U\left(i\right),U\left(t\right))={\left|\right|U\left(i\right)-U\left(i\right)\left|\right|}_2,\mathrm{\&alpha;}\&Element;\left[\mathrm{0,1}\right]\end{array}\right.$

α is weighting factor.Weighting selection criterion D (U (i), U (t)) directly reflects the similarity of U (i) and U (t); If two information vectors the closer to, then d is less, and cos β is also larger, thus D (U (i), U (t)) is also larger.Like this, in existing data message, the k group data selecting D () value maximum, by descending sort, learning of structure collection:

{(U(1),η(1)),…,(U(k),η(k))}

D(U(1),U(t))>…>D(U(k),U(t)).。

4. a kind of transmission tower tower material actual strength computing method based on data mining according to claim 1, it is characterized in that: the set of factors weight in described step 3 is determined to obtain the influence degree of certain factor to shaft tower tower material actual strength degradation ratio, by probabilistic factor sharpening, actual strength degradation ratio and the influence factor collection U local linear relation of shaft tower tower material are as follows:

η＝f _θ(U)＝U ^Tθ

Wherein: θ represents set of factors weighted value, and T represents transposition.Because current and data are in different working points, the packing density meeting present operating point U (t) may be different, data amount check for modeling is also indefinite, that is: modeling neighborhood value size is variable, in order to obtain best set of factors weight vectors θ, reduce calculated amount simultaneously, the variation range k ∈ [k of neighborhood can be preset _m, k _m] (k _m<k _m), calculating the set of factors weight vectors θ of neighbour k+1 _k+1time, directly utilize the set of factors weight vectors θ of neighbour k _k, first provide an error functions,

$\underset{\mathrm{\&theta;}}{\min }J\left(\mathrm{\&theta;}\right)=\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2$

Adopt gradient descent method, calculate set of factors weight vectors as follows,

$\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)=\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}\frac{1}{2}\underset{i=1}{\overset{m}{\mathrm{\&Sigma;}}}{(f_{\mathrm{\&theta;}}\left(U^i\right)-{\mathrm{\&eta;}}^i)}^2=(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

${\mathrm{\&theta;}}^{k+1}={\mathrm{\&theta;}}^k-\mathrm{\&alpha;}\frac{\&PartialD;}{\&PartialD;\mathrm{\&theta;}}J\left(\mathrm{\&theta;}\right)={\mathrm{\&theta;}}_i-\mathrm{\&alpha;}(f_{\mathrm{\&theta;}}\left(U\right)-\mathrm{\&eta;})U^i$

Obtaining is the model θ of neighbour k+1 _k+1, meanwhile, what also can obtain neighbour k+1 goes a transposition error value:

$e_{k+1,j}^{\mathrm{loo}}=\mathrm{\&eta;}\left(j\right)-\mathrm{\&alpha;}{\mathrm{\&theta;}}_{k+1}^{-j},j=\mathrm{1,2}...,k+1.$

In formula: represent in k+1 group data, with removing a jth model that data gained arrives; represent actual column material strength degradation rate η (j) and model error between the predicted value obtained;

Like this, what can obtain neighbour k+1 removes a transposition error collection all side and these ?

$E^{\mathrm{loo}}(k+1)=\frac{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j{\left(e_{k+1,j}^{\mathrm{loo}}\right)}^2}{\underset{j=1}{\overset{k+1}{\mathrm{\&Sigma;}}}{w}_j}.$

In formula: weighting factor the each U (j) of direct reflection go a transposition error to E ^loo(k+1) " contribution " size.The closer to the U (j) of U (t), its " contribution " is larger, otherwise less; Now, if

E ^loo(k+1)>E ^loo(k),k+1∈[k _m,k _M].

Then think and stop model " variation " returning calculating, and with model θ _kas the best model of system current time.Otherwise, by the model adopting gradient descent method to obtain, concentrate from study and choose the information vector made new advances, continue iteration, until k=k _mtill.Like this, the quality of partial model can be judged in time, obtain the best Local Linear Model meeting current time influence factor and degradation ratio relation.

5. a kind of transmission tower tower material actual strength computing method based on data mining according to claim 1, it is characterized in that: the safety assessment in described step 4, according to the influence factor of foundation and the best Local Linear Model of degradation ratio relation, calculate the degradation ratio of tower material actual strength, the degradation ratio calculating shaft tower tower material actual strength is as follows

η(t)＝＝U ^T(t)θ _k。