CN115828163A - Method for classifying health state data of power transmission tower - Google Patents

Abstract

The invention discloses a method for classifying health state data of a power transmission tower, which comprises the following steps: firstly, collecting iron tower running state data through a sensor array of inclination angles, vibration, temperature, humidity and wind power and wind speed; secondly, according to the health state data set D of the power transmission iron tower; thirdly, constructing a health state data set D of the multi-class central segmentation hyperplane support vector machine for the power transmission iron tower ₁ Classifying; fourth step, use test set D _tes Carrying out accuracy verification on a multi-class central segmentation hyperplane support vector machine model MOD; and fifthly, substituting the samples x needing to be classified into the central segmentation hyperplane support vector machine model MOD. Has the beneficial effects that: and the data set balance is improved, so that the method is suitable for machine learning. The condition that the sample simultaneously meets two hyperplanes in the learning process is avoided. In the direction of supportIn the multi-classification problem of the measuring machine, all samples can be covered only by constructing classifiers with the same number as the number of the classes.

Description

Method for classifying health state data of power transmission tower

Technical Field

The invention relates to a data classification method, in particular to a method for classifying health state data of a power transmission tower.

Background

At present, extreme weather and severe environment are important reasons causing serious tower collapse accidents of a power transmission tower. The monitoring and early warning means for the state of the iron tower is not perfect, and the iron tower may vibrate strongly and be rusted severely especially in the weather of strong wind, strong rainfall and the like. At present, a sensor array can rapidly acquire information such as an inclination angle, vibration, temperature and humidity, wind power and wind speed of an iron tower body, and rainfall, snowfall and ponding depth information can be acquired through current meteorological information. The information is integrated and extracted to obtain a data set of the health state of the iron tower, and faults of ice coating of a lead wire, a ground wire or an insulator, settlement of a tower foundation or corrosion of a tower body and the like which possibly occur to the iron tower can be predicted by training a machine learning model through the data set of the health state of the iron tower. In the existing fault method, a multi-classification support vector machine is mainly adopted to directly classify the faults of the iron tower data set. The method has two problems, namely the number of samples with each fault in the iron tower data set is unbalanced, and particularly, the number of samples generated by normal operation of the iron tower is far larger than that of the samples generated when the iron tower has faults. This situation is not conducive to the model training process of the support vector machine. Secondly, under the traditional multi-classification support vector machine, part of data in a large amount of data has fault samples which simultaneously satisfy two different hyperplanes, and the reason for the situation is that the classification algorithm of the traditional multi-classification support vector machine cannot completely cover all samples, so the algorithm of the traditional multi-classification support vector machine needs to be improved.

Disclosure of Invention

The invention aims to solve the problems that in the prior art, the classification result of the health monitoring fault of an iron tower is not accurate enough due to the imbalance of various fault data of a health monitoring data set of the iron tower and the defects of the traditional multi-classification support vector machine algorithm, and the health state data classification method of the iron tower is provided.

The invention provides a method for classifying health state data of a power transmission tower, which comprises the following steps:

the method comprises the steps of firstly, collecting iron tower running state data through a sensor array of inclination angles, vibration, temperature, humidity and wind speed, and acquiring rainfall, snow amount and water accumulation amount data at corresponding moments through real-time weather information, wherein each group is formed by temperature x ₁ Humidity x ₂ Wind direction x ₃ Wind x ₄ Amount of rainfall and snow x ₅ X amount of accumulated water ₆ Vibration signal x ₇ And the size of the dip angle x ₈ Form a sample x _i And mark the fault type y _i Wherein the fault types include four types: iron tower icing y ₁ Conductor galloping y ₂ Column foundation settling y ₃ And accumulated water corrosion y ₄ From samples x obtained during the monitoring period _i And corresponding failure type y _i Constructing a health state data set D of the power transmission tower;

secondly, according to the health state data set D of the power transmission tower, the number of various samples in the data set is adjusted by adopting a method of synthesizing few oversampling, and the balance of the classified data set is improved, specifically as follows:

(1) Clipping the samples in the health state data set D of the power transmission tower, removing the point with the minimum effect on the classification process, and reducing the calculation amount of the subsequent process;

(2) Selecting x from a certain sample by using a K-nearest neighbor algorithm _i Nearest K homogeneous samples x ^′ _i At K homogeneous samples x ^′ _i In the random selection of M samples x _n ，n＝[1,2,…M]The K-nearest neighbor algorithm is calculating x _i Neighboring homogeneous samples x ^′ _i The distance formula satisfied is as follows:

(3) Mixing M samplesRespectively and directly original copies x _i Connecting, taking any point in the period as a new sample x _N Xi is a random number, new sample x _N The structural formula of (1) is as follows:

x _N ＝x _i +ξ(x _n -x _i )ξ∈(0,1)

repeating the process under the same condition to obtain any number of samples of corresponding categories, so that the number of the samples of each category is equal, and constructing a new health state data set D of the power transmission tower ₁ ；

Thirdly, constructing a health state data set D of the multi-class central segmentation hyperplane support vector machine for the power transmission iron tower ₁ Classifying, namely coating ice on iron towers ₁ Conductor galloping y ₂ Column foundation settling y ₃ And accumulated water corrosion y ₄ In the four types of samples, the same number of samples are respectively selected to form a training set D _tri Through training set D _tri Obtaining original hyperplane OH ₁ 、OH ₂ 、OH ₃ 、OH ₄ Calculating the original hyperplane OH ₁ And OH ₂ Central dividing plane OH of ₁₂ Original hyperplane OH ₂ And OH ₃ Central dividing plane OH of ₂₃ Original hyperplane OH ₃ And OH ₄ Central dividing plane OH of ₃₄ Original hyperplane OH ₄ And OH ₁ Central dividing plane OH of ₄₁ Dividing the center into a plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Inputting the model into an original support vector machine model MOD to obtain a multi-class central segmentation hyperplane support vector machine model MOD, which is specifically as follows:

(1) Selecting a health state data set D of the power transmission tower ₁ Mark y in ₁ 、y ₂ 、y ₃ And y ₄ Of which 70% of the samples constitute the training set D _tri12 And the remaining 30% of the samples constitute test set D _tes12 Setting y ₁ ＝+1，y ₂ ＝y ₃ ＝y ₄ = 1, constraint y _i (ωx _i + b) is greater than or equal to 1, objective function

Solving an objective function by a Lagrange multiplier method:

α＝{α ₁ ,…,α _n }

(2) Solving the partial derivatives of the target omega and the target b, converting the maximum value problem into the minimum value problem, wherein the Lagrangian multiplier needs to meet the Karush-Kuhn-Tucker condition because of the inequality relation in the constraint condition, and finally solving the equation as follows:

simultaneously, constraint conditions are required to be met:

(3) Substituting alpha obtained by finally solving the equation into the objective function to obtain the corresponding original hyperplane OH ₁ ：

ω ₁ ^T x+b ₁ ＝0

(4) Repeating the above process to obtain original hyperplane OH ₂ 、OH ₃ 、OH ₄ ；

(5) Original hyperplane OH ₁ 、OH ₂ The hyperplane normal vectors are respectively omega ₁ 、ω ₂ Will ω ₁ ，ω ₂ Unitization to obtain

When original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

when the center is divided into planes OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

when original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

when the center is divided into planes OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

(6) Finding the central split plane OH ₁₂ Original hyperplane OH ₁ 、OH ₂ Public solution X of ₀ ，X ₀ As a non-zero vector of the same feature dimension. By the original hyperplane OH ₁ ，OH ₂ To X ₀ And (3) solving:

structural central dividing plane OH ₁₂ ：

OH ₁₂ :ω ₁₂ ^T x+b ₁₂ ＝0

X ₀ Is a central division plane OH ₁₂ Substituting a particular solution of (a) into the above equation yields:

(7) Repeating the processes (5) and (6) to respectively obtain a central segmentation plane OH ₂₃ 、OH ₃₄ 、OH ₄₁ By a central dividing plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Constructing a multi-class central segmentation hyperplane support vector machine model MOD;

fourth step, use test set D _tes For in multiple categoriesCarrying out accuracy verification on the central segmentation hyperplane support vector machine model MOD, drawing a result graph, and observing a classification effect;

fifthly, substituting the samples x needing to be classified into a central segmentation hyperplane support vector machine model MOD, and enabling the samples x to pass through a central segmentation plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Judging the fault type y marked by the sample x corresponding to the result value _i ，OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Of the four results, only one is positive, and the remaining three are negative if OH ₁₂ If the result value is positive, the class of the sample x is the iron tower icing y ₁ And so on.

The invention has the beneficial effects that:

the method for classifying the health state data of the power transmission iron tower is a method for classifying the health state data of the power transmission iron tower based on a multi-class central segmentation hyperplane support vector machine, and improves and classifies a health state data set of the power transmission iron tower by synthesizing a few-class oversampling method and the multi-class central segmentation hyperplane support vector machine. The method has the following advantages:

1. and the data set balance is improved, so that the data set balance is suitable for machine learning.

2. The condition that the sample simultaneously meets two hyperplanes in the learning process is avoided.

3. On the problem of multi-classification of the support vector machine, all samples can be covered only by constructing classifiers with the same number as the number of the classes.

Drawings

Fig. 1 is a schematic flow chart of a method for classifying health state data of a power transmission tower according to the present invention.

Detailed Description

Please refer to fig. 1:

the invention provides a method for classifying health state data of a power transmission tower, which comprises the following steps:

firstly, collecting iron tower running state data through a sensor array of inclination angle, vibration, temperature, humidity and wind power and wind speed, and collecting real-time weather informationAcquiring rainfall, snow amount and water accumulation amount data at corresponding time, wherein each group consists of temperature x ₁ Humidity x ₂ Wind direction x ₃ Wind x ₄ Amount of rainfall and snow x ₅ X volume of accumulated water ₆ Vibration signal x ₇ And the size of the dip angle x ₈ Form a sample x _i And mark the fault type y _i Wherein the fault types include four types: iron tower icing y ₁ Conductor galloping y ₂ Column foundation settling y ₃ And water accumulation corrosion y ₄ From samples x obtained during the monitoring period _i And corresponding failure type y _i Constructing a health state data set D of the power transmission tower;

secondly, adjusting the number of various samples in the data set by adopting a method of synthesizing few types of oversampling according to the health state data set D of the power transmission tower, and improving the balance of the classified data set, specifically as follows:

(1) Editing samples in the health state data set D of the power transmission tower, removing the point with the minimum effect on the classification process, and reducing the calculation amount of the subsequent process;

(2) Selecting x from a certain sample by using a K-nearest neighbor algorithm _i Nearest K homogeneous samples x ^′ _i At K homogeneous samples x ^′ _i In the random selection of M samples x _n ，n＝[1,2,…M]The K-nearest neighbor algorithm is calculating x _i Neighboring homogeneous samples x ^′ _i The distance formula satisfied is as follows:

(3) Respectively adding M samples to original samples x _i Concatenating, taking any point in the period as a new sample x _N Xi is a random number, new sample x _N The structural formula of (1) is as follows:

x _N ＝x _i +ξ(x _n -x _i )ξ∈(0,1)

repeating the process under the same condition to obtain samples of any number corresponding to the categories, so that the samples of all categories are equal in number and form a structureEstablishing new health state data set D of transmission tower ₁ ；

Thirdly, constructing a multi-class central segmentation hyperplane support vector machine pair power transmission tower health state data set D ₁ Classifying, including icing on iron tower ₁ Conductor galloping y ₂ Column foundation settling y ₃ And accumulated water corrosion y ₄ In the four types of samples, the same number of samples are respectively selected to form a training set D _tri Through training set D _tri Obtaining original hyperplane OH ₁ 、OH ₂ 、OH ₃ 、OH ₄ Calculating the original hyperplane OH ₁ And OH ₂ Central dividing plane OH of ₁₂ Original hyperplane OH ₂ And OH ₃ Central dividing plane OH of ₂₃ Original hyperplane OH ₃ And OH ₄ Central dividing plane OH of ₃₄ Original hyperplane OH ₄ And OH ₁ Central dividing plane OH of ₄₁ Dividing the center into planes OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Inputting the model MOD into an original support vector machine to obtain a multi-class central segmentation hyperplane support vector machine MOD, which specifically comprises the following steps:

(1) Selecting a health state data set D of the power transmission iron tower ₁ Mark y in ₁ 、y ₂ 、y ₃ And y ₄ Of which 70% of the samples constitute the training set D _tri12 And the remaining 30% of the samples constitute test set D _tes12 Setting y ₁ ＝+1，y ₂ ＝y ₃ ＝y ₄ = 1, constraint y _i (ωx _i + b) is greater than or equal to 1, objective function

Solving an objective function by a Lagrange multiplier method:

α＝{α ₁ ,…,α _n }

(2) Solving the partial derivatives of the target omega and b, converting the maximum value problem into the minimum value problem, and solving the equation finally as follows, wherein the Lagrange multiplier needs to meet the Karush-Kuhn-Tucker condition because of the inequality relation in the constraint condition:

simultaneously, constraint conditions are required to be met:

(3) Substituting alpha obtained by finally solving the equation into the objective function to obtain the corresponding original hyperplane OH ₁ ：

ω ₁ ^T x+b ₁ ＝0

(4) Repeating the above process to obtain original hyperplane OH ₂ 、OH ₃ 、OH ₄ ；

(5) Original hyperplane OH ₁ 、OH ₂ The hyperplane normal vectors are respectively omega ₁ 、ω ₂ Will be ω ₁ ，ω ₂ Unitization to obtain

When original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

while the central dividing plane OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

when original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

at the time, the center is divided intoSecant plane OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

(6) Finding the central split plane OH ₁₂ Original hyperplane OH ₁ 、OH ₂ Public solution X of ₀ ，X ₀ Is a non-zero vector of the same dimension as the feature. By the original hyperplane OH ₁ ，OH ₂ To X ₀ And (3) solving:

structural central dividing plane OH ₁₂ ：

OH ₁₂ :ω ₁₂ ^T x+b ₁₂ ＝0

X ₀ Is a central division plane OH ₁₂ Substituting a particular solution of (a) into the above equation yields:

(7) Repeating the processes (5) and (6) to respectively obtain a central segmentation plane OH ₂₃ 、OH ₃₄ 、OH ₄₁ By a central dividing plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Constructing a multi-class central segmentation hyperplane support vector machine model MOD;

fourth step, use test set D _tes Carrying out accuracy verification on the multi-class central segmentation hyperplane support vector machine model MOD, drawing a result graph, and observing a classification effect;

fifthly, substituting the samples x needing to be classified into a central segmentation hyperplane support vector machine model MOD, and enabling the samples x to pass through a central segmentation plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Judging the fault type y marked by the sample x corresponding to the result value _i ，OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Of the four results, only one was positive, the remaining three were negative, if OH ₁₂ If the result value is positive, the class of the sample x is the iron tower icing y ₁ And so on.

Claims (1)

1. A method for classifying health state data of a power transmission tower is characterized by comprising the following steps: the method comprises the following steps:

the method comprises the steps of collecting iron tower running state data through a sensor array of inclination angles, vibration, temperature, humidity and wind power and wind speed, and acquiring rainfall, snow amount and water accumulation amount data at corresponding moments through real-time weather information, wherein each group is formed by temperature x ₁ Humidity x ₂ Wind direction x ₃ Wind x ₄ Amount of rainfall and snow x ₅ X volume of accumulated water ₆ Vibration signal x ₇ And the size of the dip angle x ₈ Form a sample x _i And mark the fault type y _i Wherein the fault types include four types: iron tower icing y ₁ Conductor galloping y ₂ Column foundation settling y ₃ And accumulated water corrosion y ₄ From samples x obtained during the monitoring period _i And corresponding failure type y _i Constructing a health state data set D of the power transmission tower;

secondly, adjusting the number of various samples in the data set by adopting a method of synthesizing few types of oversampling according to the health state data set D of the power transmission tower, and improving the balance of the classified data set, specifically as follows:

(1) Editing samples in the health state data set D of the power transmission tower, removing the point with the minimum effect on the classification process, and reducing the calculation amount of the subsequent process;

(2) Selecting x from a certain sample by using a K-nearest neighbor algorithm _i Nearest K samples of the same type x' _i At K samples of the same type x' _i In the random selection of M samples x _n ，n＝[1，2，...M]K-neighbor algorithm calculates xi-adjacent homogeneous sample x' _i The distance formula satisfied is as follows:

(3) Respectively adding M samples to original samples x _i Connecting, taking any point in the period as a new sample x _N Xi is a random number, new sample x _N The structural formula of (1) is as follows:

x _N ＝x _i +ξ(x _n -x _i )ξ∈(0，1)

repeating the process under the same condition to obtain any number of samples of corresponding categories, so that the number of the samples of each category is equal, and constructing a new health state data set D of the transmission tower ₁ ；

Thirdly, constructing a health state data set D of the multi-class central segmentation hyperplane support vector machine for the power transmission iron tower ₁ Classifying, including icing on iron tower ₁ Lead waving y ₂ Column foundation settling y ₃ And water accumulation corrosion y ₄ In the four types of samples, the same number of samples are respectively selected to form a training set D _tri Through training set D _tri Obtaining original hyperplane OH ₁ 、OH ₂ 、OH ₃ 、OH ₄ Calculating the original hyperplane OH ₁ And OH ₂ Central dividing plane OH of ₁₂ Original hyperplane OH ₂ And OH ₃ Central dividing plane OH of ₂₃ Original hyperplane OH ₃ And OH ₄ Central dividing plane OH of ₃₄ Original hyperplane OH ₄ And OH ₁ Central dividing plane OH of ₄₁ Dividing the center into planes OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Inputting the model into an original support vector machine model MOD to obtain a multi-class central segmentation hyperplane support vector machine model MOD, which is specifically as follows:

(1) Selecting a health state data set D of the power transmission tower ₁ Mark y in ₁ 、y ₂ 、y ₃ And y ₄ Of which 70% of the samples constitute the training set D _tri12 And the remaining 30% of the samples constitute test set D _tes12 Setting y ₁ ＝+1，y ₂ ＝y ₃ ＝y ₄ = 1, constraint y _i (ωx _i + b) is greater than or equal to 1, objective function

Solving an objective function by a Lagrange multiplier method:

(2) Solving the partial derivatives of the target omega and the target b, converting the maximum value problem into the minimum value problem, wherein the Lagrangian multiplier needs to meet the Karush-Kuhn-Tucker condition because of the inequality relation in the constraint condition, and finally solving the equation as follows:

simultaneously, constraint conditions are required to be met:

(3) Substituting alpha obtained by finally solving the equation into the objective function to obtain the corresponding original hyperplane OH ₁ ：

ω ₁ ^T x+b ₁ ＝0

(4) Repeating the above process to obtain original hyperplane OH ₂ 、OH ₃ 、OH ₄ ；

(5) Original hyperplane OH ₁ 、OH ₂ The hyperplane normal vectors are respectively omega ₁ 、ω ₂ Will ω ₁ ，ω ₂ Unitization to obtain

When original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

when the center is divided into planes OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

when original hyperplane OH ₁ 、OH ₂ The included angle theta satisfies:

when the center is divided into planes OH ₁₂ Normal vector omega of ₁₂ Comprises the following steps:

(6) Finding the central split plane OH ₁₂ Original hyperplane OH ₁ 、OH ₂ Public solution X of ₀ ，X ₀ Is a non-zero vector with the same characteristic dimension and passes through the original hyperplane OH ₁ ，OH ₂ To X ₀ And (3) solving:

structural central dividing plane OH ₁₂ ：

OH ₁₂ ：ω ₁₂ ^T x+b ₁₂ ＝0

X ₀ Is a central division plane OH ₁₂ Substituting a particular solution of (a) into the above equation yields:

(7) Repeating the processes (5) and (6) to respectively obtain a central segmentation plane OH ₂₃ 、OH ₃₄ 、OH ₄₁ By a central dividing plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Constructing a multi-class central segmentation hyperplane support vector machine model MOD;

fourth step, use test set D _tes Carrying out accuracy verification on the multi-class central segmentation hyperplane support vector machine model MOD, drawing a result graph, and observing a classification effect;

fifthly, substituting the samples x needing to be classified into a central segmentation hyperplane support vector machine model MOD, and enabling the samples x to pass through a central segmentation plane OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Judging the fault type y marked by the sample x corresponding to the result value _i ，OH ₁₂ 、OH ₂₃ 、OH ₃₄ 、OH ₄₁ Of the four results, only one was positive, the remaining three were negative, if OH ₁₂ If the result value is positive, the class of the sample x is the iron tower icing y ₁ And so on.

Images