CN103440495A - Method for automatically identifying hydrophobic grades of composite insulators - Google Patents

Abstract

The invention belongs to the technical field of detection of performance of electric transmission line insulators, and discloses a method for automatically identifying hydrophobic grades of composite insulators. The method includes the steps of image enhancement, image filtering, image segmentation, image feature quantity extraction, neural network identification model establishment and the like. By analyzing hydrophobic images of the composite insulators, automatic identification of the hydrophobic grades of the composite insulators is achieved. Human factor influences are eliminated, the method is high in accuracy of the judgment result, easy to operate and capable of automatically identifying the hydrophobic grades of the insulators different in voltage grade, type, manufacturer, foul degree grade, operation age limit and the like.

Description

A kind of composite insulator hydrophobic grade automatic identifying method

Technical field

The invention belongs to electric transmission line isolator Performance Detection technical field, particularly a kind of composite insulator hydrophobic grade automatic identifying method.

Background technology

The silicon rubber compound insulator anti-soil good with it dodges performance and is widely used in the transmission line of electricity external insulation, and its good anti-soil dodges hydrophobic nature and the migration of hydrophobicity that performance mainly has benefited from itself.But practical operating experiences shows, due to the impact that is moved the factors such as the time limit, electric field, ultraviolet ray, filth, composite insulator silicon rubber cluster parachute hydrophobic nature will progressively descend, and even completely lose, and cause the anti-soil of composite insulator to dodge hydraulic performance decline, the generation of transmission line of electricity pollution flashover accident.For reducing pollution flashover accident, occur, carry out the detection analysis for Hydrophobicity of Composite Insulator imperative, traditional detection method adopts the water spray stage division to be sprayed water to the composite insulator of extracing on transmission line of electricity, then by the difference on comparison insulator surface water droplet attachment state and standard diagram, obtain the hydrophobic grade of insulator, this method often has subjectivity, is unfavorable for fast, accurately obtaining the hydrophobic state of composite insulator.For this reason, realize the automatic identification of composite insulator hydrophobic grade by associated picture analytical technology and artificial intelligence technology, this is for accurately knowing Hydrophobicity of Composite Insulator and ensureing that the transmission line of electricity safe and reliable operation is significant.

For objective analysis compound inslation sub-surface hydrophobic grade, eliminate the impact of subjective factor in hydrophobic classification process, the related researcher has carried out correlative study for composite insulator hydrophobic grade automatic identification technology both at home and abroad, as improved Shape Factor Method, SVM decision tree etc., carries out the Hydrophobicity of Composite Insulator detection.Result shows that these methods overcome the impact that subjective factor is judged the hydrophobic grade to a certain extent, but also there is certain deficiency, as improve Shape Factor Method and adopt the Area Ratio of the form factor of the maximum globule and the maximum globule and entire image as characteristic quantity, obtain the hydrophobic grade of composite insulator by the quantitative relationship of comparing these two characteristic quantities and hydrophobic grade, but between these two characteristic quantities and hydrophobicity rank and nonlinear relationship, the quantitative relationship provided according to the method judges that often there is error in the insulator hydrophobic state; Composite insulator hydrophobic image varies, and the SVM decision tree is difficult to carry out for extensive sample training.By setting up the automatic identifying method of hydrophobic image many reference amounts, will contribute to obtain more accurately the hydrophobic state of composite insulator, this is for guaranteeing that the transmission line of electricity safe and reliable operation is significant for this reason.

Summary of the invention

The objective of the invention is to propose a kind of automatic identifying method of composite insulator hydrophobic grade, it is characterized in that, by composite insulator hydrophobic image is analyzed, the automatic identification of realization to composite insulator hydrophobic grade, described composite insulator hydrophobic grade automatic identifying method comprises that step is as follows:

1) choose one, operation insulator for many years, gather insulator cluster parachute surface water and drip the original image adhered to, and adopt the self-adapting histogram equilibrium algorithm to realize the figure image intensifying, local histogram by computed image, then adaptive it is revised, be the unbalanced histogram transformation of original image equally distributed form, increased the dynamic range of gray-scale value, thereby reach the effect that improves the integral image contrast;

2) image filtering adopts morphology to open and close the filtering algorithm realization, and this algorithm is the structural element B that uses respectively two different scales ₁and B ₂image is carried out to the order opening and closing operation, the small-scale structure element B ₁remove the noise ability a little less than, but can keep more image edge details were information; The large-scale structure element B ₂remove the noise ability stronger, but can be fuzzy fall a lot of detailed information, therefore, two morphological structuring elements of utilization of order are opened and closed filtering to image and can either be removed noise, can keep detailed information again, reach the purpose that smoothed image is removed noise, B ₁be 3 * 3 criss-cross structural elements, B ₂be 5 * 5 diamond structure elements,

$B_1=\left(\begin{array}{ccc}0& 1& 0\\ 1& 1& 1\\ 0& 1& 0\end{array}\right)$ $B_2=\left(\begin{array}{ccccc}0& 0& 1& 0& 0\\ 0& 1& 1& 1& 0\\ 1& 1& 1& 1& 1\\ 0& 1& 1& 1& 0\\ 0& 0& 1& 0& 0\end{array}\right);$

3) image is cut apart the realization of employing Ostu threshold method, the method is that a kind of adaptive threshold of deriving on the basis of principle of least square method is determined method, the optimal threshold that utilizes the method to obtain can be divided into the pixel of entire image prospect and background two parts, make the difference maximum of prospect and background, thereby be partitioned into accurately the globule or water mark in the hydrophobic nature image;

If image pixel number is N, tonal range is [0, L-1], and the pixel count of corresponding grey scale level i is n _i, its probability is:

p _i=n _i/N(i=0,1,2...L-1) (1)

Pixel in image is divided into to two class C with threshold value T ₀and C ₁, C ₀pixel by gray-scale value between [0, T] forms, C ₁pixel by gray-scale value between [T+1, L-1] forms, and for the intensity profile probability, the average of entire image is:

${\mathrm{\μ}}_T=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ip}}_i---\left(2\right)$

C ₀and C ₁average be:

${\mathrm{\μ}}_0=\underset{i=0}{\overset{T}{\mathrm{\Σ}}}{\mathrm{ip}}_i/w_0---\left(3\right)$

${\mathrm{\μ}}_1=\underset{i=T+1}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ip}}_i/w_1---\left(4\right)$

W wherein ₀and w ₁expression formula be:

$w_0=\underset{i=0}{\overset{T}{\mathrm{\Σ}}}{p}_i---\left(5\right)$

$w_1=\underset{i=T+1}{\overset{L-1}{\mathrm{\Σ}}}{p}_i=1-w_0---\left(6\right)$

By formula (2) (3) (4) (5) (6), can be obtained:

μ _T=w ₀μ ₀+w ₁μ ₁ (7)

Inter-class variance is defined as:

${\mathrm{\σ}}_B^2=w_0{({\mathrm{\μ}}_0-{\mathrm{\μ}}_T)}^2+w_1{({\mathrm{\μ}}_1-{\mathrm{\μ}}_T)}^2---\left(8\right)$

Make T value successively in [0, L-1] scope, make

maximum T value is optimal threshold;

4) image feature amount, the form factor f of employing globule coverage rate A, the maximum globule or water mark _c, the maximum globule or the eccentric ratio e of the Area Ratio K of water mark and entire image, the maximum globule or water mark and the rectangular degree R of the maximum globule or water mark, the value that experimental results show that these five characteristic quantities is not subject to the beveled structure of insulator, the impact of filthy illumination condition, shooting distance and the shooting angle factor that distributes, takes, only, with the hydrophobic rank correlation of insulator, their expression formula is as follows:

(1) globule coverage rate

Wherein, N is the globule or the water mark number identified, S _ibe the area of i the globule or water mark, XY means the area of hydrophobic nature image;

(2) form factor

Wherein, S _maxfor the area of the maximum globule or water mark, the girth that l is the maximum globule or water mark;

(3) maximum area ratio

Wherein, S _maxfor the area of the maximum globule or water mark, the area that XY is the hydrophobic nature image.

(4) excentricity

Wherein, the major axis of the boundary rectangle that a is the maximum globule or water mark, the minor axis of the boundary rectangle that b is the maximum globule or water mark;

(5) rectangular degree

Wherein, S _maxfor the area of the maximum globule or water mark, S _mERarea for the minimum boundary rectangle of the maximum globule or water mark;

5) design of LVQ neural network recognization model

The LVQ neural network is a kind of input feedforward neural network for the supervised learning method of training competition layer, by 3 layers of neuron, formed, be input layer, competition layer and linear output layer, network structure is simple, only the interaction by internal element just can complete very complicated classification and processes, and it does not need input vector is carried out to normalization, orthogonalization process, only needs the directly distance between calculating input vector and competition layer, thereby implementation pattern identification is therefore simple.

(1) network structure determines

Using the eigenwert of image as input vector, using the hydrophobic grade of expectation as output vector.Due to 5 altogether of the image feature values extracted, so network input layer number is defined as 5, and input vector is $X={\left(\begin{array}{ccccc}A& f_c& K& e& R\end{array}\right)}^T;$ Be output as 7 hydrophobic grades, so network output layer neuron number is defined as 7, utilizes respectively 1,2,3,4,5,6,7 to mean 7 different hydrophobic grades; According to the Kolmogorve theorem, competition layer neuron number m and input layer number n have following approximation relation:

m=2n+1 （9）

Can determine that according to formula (9) network competition layer neuron number is 11.

To sum up gained, design the LVQ neural network of three layers, and structure is 5-11-7, and input layer has 5 neurons, and competition layer has 11 neurons, and output layer has 7 neurons.

(2) training of network, after network structure is determined, then the training sample data are input in the LVQ neural network designed, finally adopt the LVQ1 algorithm to be trained network, the network model trained can be identified the hydrophobic grade of composite insulator.

Described " the LVQ1 neural network algorithm is a kind of learning algorithm of competition layer being trained having under teacher's state; the basic thought of this algorithm is: calculate the nearest competition layer neuron of distance input vector; be attached thereto thereby find the linear output layer neuron connect; if the classification of input vector is consistent with the corresponding classification of linear output layer neuron, corresponding competition layer neuron weights move along the direction of input vector; Otherwise, if both classifications are inconsistent, corresponding competition layer neuron weights move along the opposite direction of input vector.Algorithmic procedure is as follows:

(1) the weights ω between initialization input layer and output layer _ij, initial learn speed is made as that η (0) (η>0), error precision are made as g, maximum frequency of training is made as t _m.

(2) by input vector $X={\left(\begin{array}{cccc}A& f_c& K& e\end{array}\right)}^T$ Be sent to input layer, and calculate the distance of competition layer neuron and input vector according to formula (10):

$\begin{array}{cc}{d}_i=\sqrt{\underset{j=1}{\overset{R}{\mathrm{\Σ}}}{(x_j-{\mathrm{\ω}}_{\mathrm{ij}})}^2}& i=\mathrm{1,2},...S\end{array}---\left(10\right)$

In formula, ω _ijfor the weights between input layer j and competition layer neuron i, S is the competition layer neuron number.

(3) select with input vector apart from minimum competition layer neuron, if d _iminimum, the neuronic class label of linear output layer that note is attached thereto is C _i.

(4) whether the judgement classification is correct, and class label corresponding to note input vector is C _xif classification is correct, i.e. C _i=C _x, according to formula (11), adjust weights; Otherwise, carry out right value update according to formula (12):

ω _{ij_new}=ω _{ij_old}+η(x-ω _{ij_old}) （11）

ω _{ij_new}=ω _{ij_old}-η(x-ω _{ij_old}) （12）

(5) renewal learning speed,

$\mathrm{\η}\left(t\right)=\mathrm{\η}\left(0\right)(1-\frac{t}{t_m})---\left(13\right)$

In formula (13), t means the t time network training; Learning rate when η (t) means the t time network training;

(6) forward step (2) to, repeat each step, until reach default frequency of training t _mor error precision g.The expression formula of error precision g is:

$g=\frac{\underset{k=1}{\overset{N}{\mathrm{\Σ}}}{(t_k-a_k)}^2}{N}---\left(14\right)$

In formula (14), N is the output layer neuron number; a _kbe k the neuronic real output value of output layer; t _kbe k the neuronic desired value of output layer.

After according to above step, that the LVQ neural metwork training is good, just can be for the identification of insulator hydrophobic grade.

The invention has the beneficial effects as follows by composite insulator hydrophobic image is analyzed, the automatic identification of realization to composite insulator hydrophobic grade, the impact of elimination human factor, the judged result accuracy is high, simple to operate, possesses the automatic recognition capability to insulator hydrophobic grades such as different electric pressures, model, producer, filthy degree grade, the operation time limits.

The accompanying drawing explanation

Fig. 1 is the enhancing design sketch that utilizes histogram equalization and self-adapting histogram equilibrium to be strengthened image respectively, a original image wherein, b. histogram equalization image, c self-adapting histogram equilibrium image.

Fig. 2 is for to utilize medium filtering, Wiener filtering and morphologic filtering image to be carried out to the noise reduction figure of level and smooth noise reduction respectively, wherein, and a noise original image, the b. dimension image of making an uproar of accepting the enemy's surrender, c. medium filtering noise reduction image, d. morphologic filtering noise reduction image.

Fig. 3 is for utilizing respectively region-growing method, the process of iteration Ostu threshold method segmentation effect figure to Image Segmentation Using, a. original image, b. region-growing method, c. process of iteration .d.Ostu threshold method.

Embodiment

The present invention proposes a kind of automatic identifying method of composite insulator hydrophobic grade, by composite insulator hydrophobic image is analyzed, set up the neural network recognization model, realize the automatic identification to composite insulator hydrophobic grade, below in conjunction with drawings and Examples, the present invention is explained.

Case 1

Choose one, operation insulator for many years, utilize respectively histogram equalization and self-adapting histogram equilibrium to be strengthened image, strengthen effect as shown in Figure 1.A original image from Fig. 1, b. histogram equalization image, c self-adapting histogram equilibrium image ratio, can find out, the picture contrast that the c self-adapting histogram equilibrium was processed obviously improves, a lot of detailed information performances also more clear; Although the picture contrast that b. histogram equalization is processed is also well a lot of than a original image, the local enhancing phenomenon of crossing has appearred, the edge of some globule is darker, makes marginal information fuzzy.So it is comparatively rational adopting self-adapting histogram equilibrium to strengthen processing to image.

Case 2

Choose one, operation insulator for many years, utilize respectively medium filtering, Wiener filtering and morphologic filtering to carry out level and smooth noise reduction to image, noise reduction as shown in Figure 2.A noise original image from Fig. 2, b. ties up the image of making an uproar of accepting the enemy's surrender, c. medium filtering noise reduction image, d. morphologic filtering noise reduction image.Can find out, the accept the enemy's surrender effect of making an uproar of b. dimension is the poorest, after noise reduction, also has a lot of noises to exist; Although c. medium filtering is all eliminated noise, topography is comparatively fuzzy; D. morphologic filtering is not only all eliminated noise, and integral image is also comparatively clear.Therefore select morphology can accurately obtain the original image of composite insulator as wave filter.

Case 3

Choose one, operation insulator for many years, utilize respectively region-growing method, process of iteration Ostu threshold method to Image Segmentation Using, segmentation effect as shown in Figure 3.A. original image from Fig. 3, b. region-growing method, c. process of iteration, d.Ostu threshold method; Can find out, the segmentation effect of d.Ostu threshold method is best, substantially there is no adhesion between the globule, and there is no the edge deficient phenomena, has effectively obtained the true picture of each water droplet, respond well.

Case 4

Choose one, operation insulator for many years, it is HC3 that the similarity of the rear careful contrast cluster parachute surface water droplet attachment state of water spray and standard diagram obtains its hydrophobic grade.By figure image intensifying, filtering, image cut apart and Characteristic Extraction after obtain the form factor f of globule coverage rate A, the maximum globule or water mark _c, the maximum globule or the eccentric ratio e of the Area Ratio K of water mark and entire image, the maximum globule or water mark and these five characteristic quantities of rectangular degree R of the maximum globule or water mark, as the form factor f with globule coverage rate A, the maximum globule or water mark _cwith these three characteristic quantities of Area Ratio K of the maximum globule or water mark and entire image, during as input vector, the hydrophobic grade that LVQ neural network recognization Model Identification goes out is HC5; As the form factor f with globule coverage rate A, the maximum globule or water mark _c, the Area Ratio K of the maximum globule or water mark and entire image and the maximum globule or water mark these four characteristic quantities of eccentric ratio e during as input vector, the hydrophobic grade that LVQ neural network recognization Model Identification goes out is HC2; Form factor f when characteristic quantity selection globule coverage rate A, the maximum globule or water mark _c, the eccentric ratio e of the maximum globule or the Area Ratio K of water mark and entire image, the maximum globule or water mark and the maximum globule or water mark these five characteristic quantities of rectangular degree R during as input vector, the hydrophobicity rank that LVQ neural network recognization Model Identification goes out is HC3.Visible, the present invention selects 5 characteristic quantities can obtain more accurately the hydrophobic grade of composite insulator as the proper vector of identification.

Case 5

Choose the insulator sample that three operation time limits are respectively 3 years, 5 years and 10 years, drip attachment state and determine that its hydrophobic grade is HC2, HC3 and HC5 by spraying water and comparing insulator cluster parachute surface water, by figure image intensifying, filtering, image, cut apart and the hydrophobic grade of these three insulators that Characteristic Extraction and LVQ neural network recognization model obtain is also HC2, HC3 and HC5, match with actual hydrophobicity rank, show that the method accurately and reliably.

Case 6

Choose 4, the insulator sample that gradation for surface pollution is respectively I, II, III and IV level, drip attachment state and determine that its hydrophobic grade is respectively HC2, HC4, HC3 and HC2 by spraying water and comparing insulator cluster parachute surface water, be respectively HC2, HC4, HC3 and HC2 by the hydrophobicity rank that figure image intensifying, filtering, image are cut apart and Characteristic Extraction and LVQ neural network recognization model obtain these four insulators, match with actual hydrophobicity rank, show that the method to the identification of defilement and insulation sub-surface hydrophobic state accurately and reliably.

Claims (2)

1. the automatic identifying method of a composite insulator hydrophobic grade, it is characterized in that, by composite insulator hydrophobic image is analyzed, realize the automatic identification to composite insulator hydrophobic grade, described composite insulator hydrophobic grade automatic identifying method comprises that step is as follows:

1) choose one, operation insulator for many years, gather insulator cluster parachute surface water and drip the original image adhered to, and adopt the self-adapting histogram equilibrium algorithm to realize the figure image intensifying, local histogram by computed image, then adaptive it is revised, be the unbalanced histogram transformation of original image equally distributed form, increased the dynamic range of gray-scale value, thereby reach the effect that improves the integral image contrast;

2) image filtering adopts morphology to open and close the filtering algorithm realization, and this algorithm is the structural element B that uses respectively two different scales ₁and B ₂image is carried out to the order opening and closing operation, the small-scale structure element B ₁remove the noise ability a little less than, but can keep more image edge details were information; The large-scale structure element B ₂remove the noise ability stronger, but can be fuzzy fall a lot of detailed information, therefore, two morphological structuring elements of utilization of order are opened and closed filtering to image and can either be removed noise, can keep detailed information again, reach the purpose that smoothed image is removed noise, B ₁be 3 * 3 criss-cross structural elements, B ₂be 5 * 5 diamond structure elements,

$B_1=\left(\begin{array}{ccc}0& 1& 0\\ 1& 1& 1\\ 0& 1& 0\end{array}\right)$ $B_2=\left(\begin{array}{ccccc}0& 0& 1& 0& 0\\ 0& 1& 1& 1& 0\\ 1& 1& 1& 1& 1\\ 0& 1& 1& 1& 0\\ 0& 0& 1& 0& 0\end{array}\right);$

3) image is cut apart the realization of employing Ostu threshold method, the method is that a kind of adaptive threshold of deriving on the basis of principle of least square method is determined method, the optimal threshold that utilizes the method to obtain can be divided into the pixel of entire image prospect and background two parts, make the difference maximum of prospect and background, thereby be partitioned into accurately the globule or water mark in the hydrophobic nature image;

If image pixel number is N, tonal range is [0, L-1], and the pixel count of corresponding grey scale level i is n _i, its probability is:

p _i=n _i/N(i=0,1,2...L-1) (1)

Pixel in image is divided into to two class C with threshold value T ₀and C ₁, C ₀pixel by gray-scale value between [0, T] forms, C ₁pixel by gray-scale value between [T+1, L-1] forms, and for the intensity profile probability, the average of entire image is:

${\mathrm{\μ}}_T=\underset{i=0}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ip}}_i---\left(2\right)$

C ₀and C ₁average be:

${\mathrm{\μ}}_0=\underset{i=0}{\overset{T}{\mathrm{\Σ}}}{\mathrm{ip}}_i/w_0---\left(3\right)$

${\mathrm{\μ}}_1=\underset{i=T+1}{\overset{L-1}{\mathrm{\Σ}}}{\mathrm{ip}}_i/w_1---\left(4\right)$

W wherein ₀and w ₁expression formula be:

$w_0=\underset{i=0}{\overset{T}{\mathrm{\Σ}}}{p}_i---\left(5\right)$

$w_1=\underset{i=T+1}{\overset{L-1}{\mathrm{\Σ}}}{p}_i=1-w_0---\left(6\right)$

By formula (2) (3) (4) (5) (6), can be obtained:

μ _T=w ₀μ ₀+w ₁μ ₁ (7)

Inter-class variance is defined as:

${\mathrm{\σ}}_B^2=w_0{({\mathrm{\μ}}_0-{\mathrm{\μ}}_T)}^2+w_1{({\mathrm{\μ}}_1-{\mathrm{\μ}}_T)}^2---\left(8\right)$

Make T value successively in [0, L-1] scope, make

maximum T value is optimal threshold;

4) image feature amount, the form factor f of employing globule coverage rate A, the maximum globule or water mark _c, the maximum globule or the eccentric ratio e of the Area Ratio K of water mark and entire image, the maximum globule or water mark and the rectangular degree R of the maximum globule or water mark, the value that experimental results show that these five characteristic quantities is not subject to the beveled structure of insulator, the impact of filthy illumination condition, shooting distance and the shooting angle factor that distributes, takes, only, with the hydrophobic rank correlation of insulator, their expression formula is as follows:

(1) globule coverage rate

Wherein, N is the globule or the water mark number identified, S _ibe the area of i the globule or water mark, XY means the area of hydrophobic nature image;

(2) form factor

Wherein, S _maxfor the area of the maximum globule or water mark, the girth that l is the maximum globule or water mark;

(3) maximum area ratio

Wherein, S _maxfor the area of the maximum globule or water mark, the area that XY is the hydrophobic nature image;

(4) excentricity

Wherein, the major axis of the boundary rectangle that a is the maximum globule or water mark, the minor axis of the boundary rectangle that b is the maximum globule or water mark;

(5) rectangular degree

Wherein, S _maxfor the area of the maximum globule or water mark, S _mERarea for the minimum boundary rectangle of the maximum globule or water mark;

5) design of LVQ neural network recognization model

(1) network structure determines

Using the eigenwert of image as input vector, using the hydrophobic grade of expectation as output vector, due to 5 altogether of the image feature values extracted, so network input layer number is defined as 5, and input vector is $X={\left(\begin{array}{ccccc}A& f_c& K& e& R\end{array}\right)}^T;$ Be output as 7 hydrophobic grades, so network output layer neuron number is defined as 7, utilizes respectively 1,2,3,4,5,6,7 to mean 7 different hydrophobic grades; According to the Kolmogorve theorem, competition layer neuron number m and input layer number n have following approximation relation:

m=2n+1 （9）

Can determine that according to formula (9) network competition layer neuron number is 11.

To sum up gained, design the LVQ neural network of three layers, and structure is 5-11-7, and input layer has 5 neurons, and competition layer has 11 neurons, and output layer has 7 neurons;

(2) network training

Then network structure is input to the training sample data in the LVQ neural network designed after determining, finally adopts the LVQ1 algorithm to be trained network, and the network model trained can be identified the hydrophobic grade of composite insulator.

2. a kind of automatic identifying method of composite insulator hydrophobic grade according to claim 1, it is characterized in that, in described step 5), the LVQ neural network is a kind of input feedforward neural network for the supervised learning method of training competition layer, by 3 layers of neuron, formed, it is input layer, competition layer and linear output layer, network structure is simple, only the interaction by internal element just can complete very complicated classification processing, and it does not need input vector is carried out to normalization, orthogonalization process, only need directly to calculate the distance between input vector and competition layer, thereby implementation pattern identification, therefore simple.

Images