CN106595551A - Icing thickness detection method for power transmission line icing image based on deep learning - Google Patents

Abstract

The invention discloses an icing thickness detection method for a power transmission line icing image based on deep learning, belongs to the field of digital image recognition, and aims at solving problems that the sensitivity and reliability of conventional tension monitoring are not high and improving the icing thickness monitoring accuracy and automation degree of a power transmission line. The method is used for the icing thickness monitoring and over-range warning for a system power transmission line, and comprises the steps: (1), collecting an icing image; (2), carrying out the preprocessing of an image, and building a data set; (3), building a convolution neural network; (4), carrying out the training and testing of a model; (5), extracting the icing thickness information, and transmitting the icing thickness information to a control center. The method introduces a digital image feature recognition method to the icing thickness detection of a power transmission and a tower, automatically extracts the thickness information through employing the icing shape features in the image, formulates a deicing plan for an operation and maintenance worker, and provides a new visual and intelligent solution for the safe and stable operation of a power system.

Description

Based on ice covering thickness detection method in the powerline ice-covering image of deep learning

Technical field

The invention belongs to digital image understanding technical field, and in particular to cover in a kind of image based on deep learning algorithm Ice thickness condition detection method, can be used for the monitoring of power system power transmission network equipment icing and alarm of transfiniting.

Background technology

Power network safety operation is self-evident to the importance of the national economic development, deepen continuously with Power System Interconnection and The progressively enforcement of electricity market, the running environment of electrical network are also more complicated, the stability and reliability of electrical network are proposed higher Requirement.China is vast in territory, weather is various, with a varied topography, and the power network for spreading all over the country is often subject to various natural disasters Destruction, China major part Jing is often because extreme low temperature and icing cause large area blackout.Icing disaster can cause electrical network to be sent out Raw mechanical breakdown and electric fault, such as transformer station are stopped transport, shaft tower collapses, ice dodges tripping operation, line oscillation and substation equipment and damages Etc. accident.

Icing causes substantial equipment loss and causes large area blackout to China's electrical network, and ice-covering area occurs often Weather conditions are severe, and transport and communication interrupts, and repairing difficulty is big, causes large area blackout, has a strong impact on power supply reliability. The icing of transmission line of electricity and power equipment is objective reality, it is impossible to be inherently eliminated.To reduce the disaster that icing brings, must Electrical equipment icing in electrical network must be protected, eliminate icing potential safety hazard in time.

Icing protection at present is main to be passed by installing on electrical equipment by monitoring and suppressing two methods, monitoring meanss Sensor and video camera realize on-line monitoring ice coating state, or find that failure hidden to critical circuits inspection by manual inspection mode Suffer from；It is domestic often to be affected area to be mounted with the icing of ice covering monitoring system monitoring electrical network critical circuits and node by icing disaster State, but the ice damage accident for taking place frequently in recent years proves that current several monitoring meanss cannot also meet power system security, stablize Service requirement, as a example by 2014, because icing cause tripping operation 597 times, trip-out rate be 0.103 time/hundred km years, overlap into Power 46.4%, increases by 376 times compared with 2013 (221 times), and amplification is 170.1%.Icing in 2014 causes that failure is non-to stop 320 times, the non-rate of stopping of failure was 0.055 time/hundred km years, and icing causes the non-number of times that stops of failure to be about 2013 (63 within 2014 It is secondary) 5 times.Analyze current icing monitoring and be difficult to the reason for meeting operation of power networks requirement, may be summarized to be：

(1) Sensor monitoring icing is affected larger by working environment, and harsh climate meteorological condition or the interference of electromagnetic field are equal The certainty of measurement of sensor can be reduced；

(2) video camera monitoring judges current ice coating state by shooting and monitoring point shaft tower or line ice coating image, but due to Lack the effective process and Application way to icing image, it is impossible to from the reliable icing information of acquisition；

(3) manual inspection mode or helicopter routing inspection high cost, efficiency are low, it is difficult to which the whole network is monitored；

(4) system can not find icing hidden danger in time, cause ice-coating pre-warning send in time, add the phase of anti-ice operation To delayed, icing cannot be eliminated in time.

The content of the invention

In order to solve above-mentioned technical problem, the invention provides a kind of powerline ice-covering figure based on deep learning algorithm As recognition methodss, can be according to the icing image automatic identification ice covering thickness of input, it is ensured that the timely line ice coating of power department Situation.

The technical solution adopted in the present invention is：Ice covering thickness in a kind of powerline ice-covering image based on deep learning Detection method, it is characterised in that comprise the following steps：

Step 1：Obtain icing view data and corresponding answer pulling force Monitoring Data；

Step 2 carries out pretreatment to original icing image, the size of original icing image is processed into size consistent Image, with pulling force sensor measure ice covering thickness as image tag；

Step 3：Deep learning convolutional neural networks model is set up, for amount of images and size, corresponding model is set up Parameter, arranges the unit number and activation primitive per layer network；

Step 4：Adjustment weights training pattern, carry out feature extraction and combination to image, judge and export ice covering thickness；

Step 5：Analysis model trains classification results, extracts the ice covering thickness information of icing image.

China's electric icing monitoring real-time and effectiveness Shortcomings, as long as reason is to improve Sensor monitoring effect master To be improved from hardware aspect；The present invention starts with from icing view data, have studied a kind of more quick, accurate measurements icing The method of state, hidden danger of fixing a breakdown in time improve the reliability of electrical network ice covering monitoring system, ensure that power system security is steady It is fixed to run.

The present invention is incorporated into digital picture characteristic recognition method in the ice covering thickness detection of power transmission line and shaft tower, using figure The morphological characteristic of icing as in automatically extracts thickness information, is that operation maintenance personnel formulates deicing plan, is to ensure power system security Stable operation provides a kind of new means directly perceived and intelligentized.

Figure of description

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

Fig. 2 is the original icing image of the embodiment of the present invention；

Fig. 3 is the iterative method segmentation image result of the embodiment of the present invention；

Fig. 4 is the LoG operator edge detection results of the automatic threshold of the embodiment of the present invention；

Fig. 5 is the Prewitt operator edge detection results of the automatic threshold of the embodiment of the present invention；

Fig. 6 is the Sobel operator edge detection results of the automatic threshold of the embodiment of the present invention；

Fig. 7 is the convolutional neural networks feature extraction schematic diagram of the embodiment of the present invention；

Fig. 8 is the convolution process schematic diagram of the embodiment of the present invention；

Fig. 9 is the pond process schematic of the embodiment of the present invention；

Figure 10 is 1 accuracy of identification contrast of convolutional neural networks ergodic data of each structure of the embodiment of the present invention；

Figure 11 is 5 accuracy of identification contrasts of convolutional neural networks ergodic data of each structure of the embodiment of the present invention；

Figure 12 is 10 accuracy of identification contrasts of convolutional neural networks ergodic data of each structure of the embodiment of the present invention；

Figure 13 is 15 accuracy of identification contrasts of convolutional neural networks ergodic data of each structure of the embodiment of the present invention；

Figure 14 is 20 accuracy of identification contrasts of convolutional neural networks ergodic data of each structure of the embodiment of the present invention；

Figure 15 is the convolutional neural networks difference Model Identification error contrast of the embodiment of the present invention.

Specific embodiment

Understand for the ease of those of ordinary skill in the art and implement the present invention, below in conjunction with the accompanying drawings and embodiment is to this It is bright to be described in further detail, it will be appreciated that enforcement example described herein is merely to illustrate and explains the present invention, not For limiting the present invention.

Fig. 1 is asked for an interview, ice covering thickness detection in a kind of powerline ice-covering image based on deep learning that the present invention is provided Method, comprises the following steps：

Step 1：From the ice covering monitoring system of electrical network department collect icing view data and it is corresponding answer pulling force monitoring number According to；

The icing image of collection, including the icing image such as power transmission line, shaft tower, gold utensil, electrical equipment such as transformator, collection Icing image should be tried one's best clearly；Collect answer pulling force data include shaft tower model, position and answer pulling force sensor measure equivalence Ice covering thickness.

Step 2：Pretreatment is carried out to original image, picture size is adjusted to identical, icing image data set is set up；

Pretreatment, including image segmentation and edge extracting are carried out to original image；

Image segmentation is the technology and process for dividing the image into the region of each tool characteristic and extracting interesting target；

Threshold segmentation is a kind of method commonly used in image segmentation, and all gray scales are determined more than or equal to the pixel of threshold values To belong to object, gray value represents prospect with " 255 ", and otherwise these pixels are excluded beyond object area, threshold segmentation Including Two-peak method and iterative method；

Two-peak method divides the image into foreground and background two parts, and the intensity profile curve approximation of image is just considered by two State distribution functionWithBe formed by stacking, the rectangular histogram of image will occur two detached peak values, it is bimodal it Between trough at be exactly image threshold values be located；

Iterative method is the improvement to Two-peak method, selects an approximate threshold values T first, divides the image into into part R₁And R₂, meter Calculate region R₁And R₂Mean μ₁And μ₂, select new partition threshold T=(μ₁+μ₂)/2, steps be repeated alternatively until μ₁And μ₂No longer Till change, the present invention carries out image segmentation process to image using iterative method；

Marginal point in edge extracting detection image first, then edge point is connected into into profile, so as to constitute cut zone, As edge is the demarcation line of to be fetched target and background, extracting edge could separate target and background, and gradient-norm is calculated Son has the property of shift invariant and isotropism matter, it is adaptable to the side of rim detection, and the direction of grey scale change, i.e. border To then can be by θ_g=arctan (f_y/f_x) determine, wherein f_xAnd f_yIt is the direction mould of x and y respectively, θ_gIt is the inspection of consecutive image edge Survey the direction of maximum of gradients；Operator is represented by the present invention with differential operator form, then with fast convolution function realizing.

Step 3：Deep learning convolutional neural networks model is set up, for picture number and size, sets up corresponding suitable Model parameter, arranges the unit number and activation primitive per layer network；

Deep learning convolutional neural networks are set up, is a kind of deep learning model transformed based on BP neural network, its power The shared network structure of value can reduce the complexity of network model, reduce the quantity of weights, and the advantage in the input of network is What is showed during multidimensional image becomes apparent from, allow image directly as the input of network, it is to avoid multiple in tional identification algorithm Miscellaneous feature extraction and data reconstruction processes；

Deconvoluted with convolution kernel in convolutional layer the input of this layer.First by each output characteristic figure position phase of last layer Same data carry out convolution with the convolution kernel of this layer, then by same position convolution all results addeds, obtain this layer of output characteristic The output of figure correspondence position.In order to reduce number of parameters, model training difficulty is reduced, deep learning adopts weights shared mechanism. Same output characteristic figure uses same convolution kernel, and all each self-corresponding wave filter of convolution kernel is corresponding each time, one Convolution kernel only extracts a kind of feature, it is ensured that feature extraction does not occur aliasing；

After feature is obtained by convolution, the feature obtained with all extractions goes to train grader, the present invention to utilize Softmax graders, each feature and image convolution can obtain the convolution feature of a n dimension, due to there is n feature, learn Practise a grader for having more than the input of 20,000,000 features to be inconvenient, overfitting easily occur；

Step 4：Adjustment weights training pattern, carry out feature extraction and extraction to image, judge and export ice covering thickness, Compare the thickness information of image tag；

The convolutional neural networks of foundation are trained and test, and obtain meeting the model of required precision in detecting picture Ice covering thickness, its training process comprises the steps：

(1) sensitivity and error correction；

For C classification problems, N number of training sample is had, the mean square deviation of model is represented by：

WhereinThe kth class desired output of n-th sample is represented,Represent the kth class reality output of n-th sample.

For n-th sample, reality output is represented by with the mean square deviation of preferable output：

It is assumed that L is output layer, l is hidden layer, and 1 is input layer；The activation of l layers is output as：x^l=f (u^l), wherein u^l= W^lx^l-1+b^l, f () is activation primitive, W^lIt is the weights of l layers, b^lIt is the biasing of l layers；Defining sensitivity is：

WhereinTherefore the sensitivity of l layers and output layer is expressed as：

Thus obtain error correction (η is learning rate) to be represented by：

(2) propagated forward

1) convolutional layer

Assume that l layers are convolutional layers, then the characteristic pattern and characteristic pattern size of this layer of output is expressed as：

Output.size=input.size-ker nel.size+1

Wherein,It is i-th output of l-1 layers,It is l layers for j-th convolution kernel of i-th input,It is J-th biasing of l layers, f () is activation primitive,It is j-th output of l layers.

The feature extraction of convolutional neural networks has two features：

I. by convolution, represent that with a pixel of output characteristic figure the pixel of the regional area of input feature vector figure is special Levy, this is the feature extraction of convolutional neural networks, while also reducing data dimension；

II. weights are shared, and same characteristic pattern uses same convolution kernel, extract a feature, it is possible to reduce parameter number Amount, reduces time complexity.

2) sub-sampling layer

Sub-sampling layer can be regarded as pond process, and a reduction process so that input feature vector figure does not overlappingly exist Again represent on output characteristic figure, i.e. the combination of feature；Pondization can also reduce data dimension in addition, accelerate calculating speed.It is defeated The expression for going out figure is represented by：

Wherein down (x) is that the pixel region to input picture n × n carries out sampling operation,It is controlling elements, by sampling knot Fruit numerical control is in colour element numerical range, while reducing noise jamming.

(3) back propagation

1) convolutional layer

The sensitivity of convolutional layer is represented by shown in following formula, it is contemplated that be sub-sampling layer before and after convolutional layer, rewritable to be：

Wherein, β is weights,

Replace ∑ δ^l+1。

The gradient of base and convolution kernel is represented by：

Wherein u, v are deconvoluted the corresponding local of last layer image with convolution kernel.Wherein,It is j-th convolution kernel of l layers Sensitivity,It is the factor of momentum of l-1 layers；

2) sub-sampling layer

Save in propagated forward：Therefore weights gradient is：

Wherein,It is the down-sampling handling function of l layers.

(4) combinations of features

One important feature of deep learning network is automatic learning characteristic, and expression in the training process as learns spy When levying the combination of figure, for each feature extracted gives weights, propagated forward is repeated with back propagation more positive error simultaneously Adjustment weights, reach the purpose of characteristic optimization combination.Feature weight α_ijRepresent, represent wherein the i-th of j-th output characteristic figure Individual input feature vector figure weights or contribution, are generally represented with following formula：

Wherein, c_ijRepresent i-th input feature vector figure weights in j-th output characteristic figure, ∑_kexp(c_kj) represent j-th it is defeated Go out characteristic pattern all weights and.

Then j-th feature output is rewritable is：

Wherein, f () is activation primitive, α_ijThe weights of output characteristic figure are represented,It is this layer of input,It is convolution kernel,It is this layer of biasing.

Above formula meets：0≤α_ij≤1；

For single output unit, ignore footmark j, due to meeting：

With

After obtaining base gradient, convolution kernel gradient, connection weight gradient and the feature weight gradient of hidden layer, error correction is then Shown in table 1 (η is learning rate).

1 convolutional neural networks of table more on the occasion of

Wherein, η is learning rate,It is the gradient of base,It is the gradient of convolution kernel,It is to connect gradient, α_iIt is special Levy weights.

Step 5：Analysis model trains classification results, extracts the ice covering thickness information of icing image and passes back in icing monitoring The heart.

The step of the present embodiment 1, collects the original icing view data of electrical network and answers pulling force sensor data, shown in accompanying drawing 2 Be powerline ice-covering image, icing image request is clear to icing photographing section, covers without muddy thing, answers pulling force data bag Include the model of shaft tower, position and answer the ice covering thickness that pulling force sensor measures, to find the position of fault in time；

The step of the present embodiment 2, passes through image segmentation and edge extracting pretreatment icing image, by the size of original image It is set to size unanimously, with the ice covering thickness of pulling force sensor measurement as image tag；Image segmentation is carried out to image And edge treated, to remove noise, improving feature extraction efficiency, image segmentation and edge extracting are to exclude environment and unrelated Interference of the object to icing feature extraction, as ultra-high-tension power transmission line surrounding is complex, often has the environment such as trees, insecticide Factor is disturbed, and needs to reject the irrelevant factor in icing image using image Segmentation Technology, it is to avoid disturb feature extraction；It is attached Fig. 3 is iterative method segmentation image effect, and accompanying drawing 4-6 is the LoG operator edge detections of automatic threshold, Prewitt operators side respectively Edge is detected and Sobel operator edge detection results；

The step of the present embodiment 3, sets up convolutional neural networks, and convolutional neural networks utilization space relation is reduced needs study Number of parameters with before improving to, back-propagation algorithm training performance, accompanying drawing 7 is the brief description to model structure；

Convolutional neural networks include input layer, hidden layer and output layer, and the inside hidden layer of convolutional neural networks is convolution In layer and pond stacking generation, build, and convolutional layer can extract data characteristicses, also be feature extraction layer, are substantially convolution, and accompanying drawing 8 is volume Product process schematic, sub-sampling layer is also Feature Mapping layer, by pixel weighted sum, obtaining feature by activation primitive and reflecting Penetrate, accompanying drawing 9 is pond process schematic；

In convolutional neural networks, the input of the part (local experiences area) of image as the lowermost layer of hierarchical structure, Information is transferred to different layers again successively, and per layer is gone to obtain the most significant feature of observation data by a digital filter, Therefore, it is possible to obtain the marked feature of the observation data to translation, scaling and invariable rotary,

Model parameter is defined in step 3 and parameters are initialized, and sets activation primitive as " sigmoid "；

The step of the present embodiment, 4 pairs of convolutional neural networks set up were trained, and can adjust model according to training effect The relevant parameters such as structure, train epochs, batch processing size, activation primitive, per layer of neuron number, neuronal quantity；

As model parameter amount is big, the present invention is trained using 15235 icing pictures, according to electrical network to ice coating state Monitoring mechanism, is divided into without icing (0cm), slight icing (0～5cm), moderate icing (5 according to model judgment models monitoring thickness ～10cm), serious icing (10～15cm), dangerous icing (15～20cm) and fault pre-alarming icing (more than 20cm) six etc. Level, output had previously been compared judgement with image, alignment error.

The step of the present embodiment 5 analysis model training classification results, extract the ice covering thickness information of icing image and pass back Icing Surveillance center；

The present invention has built four models of convolutional neural networks, is (1) 6-4-12-2, (2) 12-4-24-2 respectively, (3) 4-2-8-2-16-2-32-2, (4) 16-2-8-2-4-2-2-2.Wherein odd-level is convolutional layer, and even level is pond layer.Four kinds In ergodic data difference number of times, each model mean square deviation changes model.The mean square deviation of model weighs the training effect of model, works as mould The mean square deviation of type is less, represents that training effect is better, and parameter setting is more reasonable, and resulting output is closer to desired result. Step number is to represent adjustment weights number of times, in deep learning algorithm, is generally trained training sample in batches, after having trained a collection of sample By weighed value adjusting once, improving training speed and training effect；

Accompanying drawing 10 represents 4 kinds of model structures：(1) 6-4-12-2, (2) 12-4-24-2, (3) 4-2-8-2-16-2-32-2, (4) 16-2-8-2-4-2-2-2, once, the situation of change of the mean square deviation of its training process, abscissa are represented ergodic data respectively Adjusting parameter number of times, vertical coordinate represent mean square deviation；Curve tendency can reflect the training effect of model, while according to training effect Adjustment model parameter, to reach optimal training effect.In the same manner, accompanying drawing 11-14 represents four kinds of model structure ergodic datas 5 respectively It is secondary, 10 times, 15 times and when 20 times, the situation of change of mean square deviation, generally traversal number of times are more, and the classification performance of model is better.

Accompanying drawing 15 is to 4 kinds of model structures：(1) 6-4-12-2, (2) 12-4-24-2, (3) 4-2-8-2-16-2-32-2, (4) relative analyses of the classification results of 16-2-8-2-4-2-2-2, abscissa are ergodic data number of times, and vertical coordinate is that classification is missed Difference.

It should be appreciated that the part that this specification is not elaborated belongs to prior art.

It should be appreciated that the above-mentioned description for preferred embodiment is more detailed, therefore can not be considered to this The restriction of invention patent protection scope, one of ordinary skill in the art are being weighed without departing from the present invention under the enlightenment of the present invention Under the protected ambit of profit requirement, replacement can also be made or deformed, be each fallen within protection scope of the present invention, this It is bright scope is claimed to be defined by claims.

Claims (10)

1. ice covering thickness detection method in a kind of powerline ice-covering image based on deep learning, it is characterised in that include with Lower step：

Step 1：Obtain icing view data and corresponding answer pulling force Monitoring Data；

Step 2 carries out pretreatment to original icing image, and the size of original icing image is processed into the consistent figure of size Picture, with the ice covering thickness of pulling force sensor measurement as image tag；

Step 3：Deep learning convolutional neural networks model is set up, for amount of images and size, corresponding model parameter is set up, Unit number and activation primitive per layer network is set；

Step 4：Adjustment weights training pattern, carry out feature extraction and combination to image, judge and export ice covering thickness；

Step 5：Analysis model trains classification results, extracts the ice covering thickness information of icing image.

2. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 1, which is special Levy and be：Icing image described in step 1, including ice covering on transmission lines image, shaft tower icing image, gold utensil icing image, electrically Equipment icing image；It is described to answer pulling force Monitoring Data to include shaft tower model, position, icing equivalent thickness.

3. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 1, which is special Levy and be：Pretreatment, including image segmentation and edge extracting are carried out to original icing image described in step 2；Described image point It is that image segmentation process is carried out to image using iterative method to cut, and in the edge extracting, gradient modules operator is with differential operator form table Show, realized with fast convolution function.

4. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 1, which is special Levy and be：Deep learning convolutional neural networks are set up described in step 3, is a kind of deep learning transformed based on BP neural network Model, the input of this layer that deconvoluted with convolution kernel in convolutional layer first will be each output characteristic figure position of last layer identical The convolution kernel of data and this layer carry out convolution, then by same position convolution all results addeds, obtain this layer of output characteristic figure The output of correspondence position.

5. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 4, which is special Levy and be：In order to reduce number of parameters, model training difficulty is reduced, using weights shared mechanism, same output characteristic figure makes Same convolution kernel is used, all each self-corresponding wave filter of convolution kernel is corresponding each time, a convolution kernel only extracts a kind of spy Levy, it is ensured that feature extraction does not occur aliasing.

6. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 1, which is special Levy and be：Convolutional neural networks described in step 4 to setting up are trained and test, and its training process includes following sub-step Suddenly：

Step 4.1：Meter sensitivity and error correction；

Step 4.2：Propagated forward；

Step 4.3：Back propagation；

Step 4.4：Combinations of features；For each feature extracted gives weights, propagated forward is repeated with back propagation more Positive error simultaneously adjusts weights, reaches the purpose of characteristic optimization combination.

7. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 6, which is special Levy and be：In step 4.1, for C classification problems, N number of training sample is had,Represent that the kth class expectation of n-th sample is defeated Go out,The kth class reality output of n-th sample is represented, then the output layer error of model is equal with reality output with preferable output Variance is expressed as：

$E^N=\frac{1}{2}\underset{n=1}{\overset{N}{\Σ}}\underset{k=1}{\overset{C}{\Σ}}{(t_k^n-y_k^n)}^2;$

For n-th sample, reality output is represented by with the mean square deviation of preferable output：

$E^n=\frac{1}{2}\underset{k=1}{\overset{C}{\Σ}}(t_k^n-y_k^n)=\frac{1}{2}\left|\right|t^n-y^n|{|}_2^2;$

It is assumed that L is output layer, l is hidden layer, and 1 is input layer；The activation of l layers is output as：x^l=f (u^l), wherein u^l=W^lx^l-1 +b^l, f () is activation primitive, W^lIt is the weights of l layers, b^lIt is the biasing of l layers；Defining sensitivity is：

$\frac{\∂E}{\∂b}=\frac{\∂E}{\∂u}\·\frac{\∂u}{\∂b}=\mathrm{\δ}$

WhereinTherefore the sensitivity of l layers and output layer is expressed as：

Thus obtaining error correction is：

$\frac{\∂E}{\∂W^l}=x^{l-1}{\left({\mathrm{\δ}}^l\right)}^T$

${\mathrm{\ΔW}}^l=-\mathrm{\η}\frac{\∂E}{\∂W^l}$

Wherein, η is learning rate.

8. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 7, which is special Levy and be：In step 4.2, for convolutional layer, it is assumed that l layers are convolutional layers, then this layer is exported characteristic pattern and characteristic pattern size Respectively：

$x_j^l=f(\underset{i{\∈}_{Mj}}{\Σ}{x}_i^{l-1}*k_{ij}^l+b_j^l)$

Output.size=input.size-ker nel.size+1

Wherein,It is i-th output of l-1 layers,It is l layers for j-th convolution kernel of i-th input,It is l layers J-th biasing, f () is activation primitive,It is j-th output of l layers；

For sub-sampling layer, output figure is：

$x_j^l=f\left({\mathrm{\β}}_j^{l}down\right(x_j^{l-1})+b_j^l)$

Wherein, down (x) is that the pixel region to input picture n × n carries out sampling operation,It is controlling elements, by sampled result number Value control is in colour element numerical range, while reducing noise jamming.

9. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 8, which is special Levy and be：In step 4.3, the sensitivity of convolutional layer can be represented：

δ^l=(W^l+1)^T(∑δ^l+1f'(u^l))；

Wherein, β is weights,Replace ∑ δ^l+1；

The gradient of base and convolution kernel is represented by：

$\frac{\∂E}{\∂b_j}={\left(\underset{u,v}{\Σ}\right({\mathrm{\δ}}_j^l\left)\right)}_{u,v};$

$\frac{\∂E}{\∂k_{ij}}=\underset{n,v}{\Σ}\left({\left({\mathrm{\δ}}_j^l\right)}_{u,v}{\left(p_i^{l-1}\right)}_{u,v}\right);$

Wherein u, v are deconvoluted the corresponding local of last layer image with convolution kernel；

For sub-sampling layer, save in propagated forwardTherefore weights gradient is：

10. the ice covering thickness detection method in the powerline ice-covering image of deep learning according to claim 9, which is special Levy and be：In step 4.4, feature weight is α_ij, represent j-th output characteristic figure wherein i-th input feature vector figure weights or Contribution；

${\mathrm{\α}}_{ij}=\frac{\exp \left(c_{ij}\right)}{{\mathrm{\Σ}}_k\exp \left(c_{kj}\right)};$

Then j-th feature output is rewritable is：

$x_j^l=f\left(\underset{i=1}{\overset{N_{in}}{\Σ}}{\mathrm{\α}}_{ij}\right(x_i^{l-1}*k_i^l)+b_j^l);$

Above formula meets：

For single output unit, ignore footmark j, due to meeting：

After obtaining base gradient, convolution kernel gradient, connection weight gradient and the feature weight gradient of hidden layer, error correction is then table Shown in 1：

1 convolutional neural networks of table more on the occasion of

Wherein, η is learning rate,It is the gradient of base,It is the gradient of convolution kernel,It is to connect gradient, α_iIt is feature power Value.