CN111914720A - Method and device for identifying insulator burst of power transmission line - Google Patents

Method and device for identifying insulator burst of power transmission line Download PDF

Info

Publication number
CN111914720A
CN111914720A CN202010732376.7A CN202010732376A CN111914720A CN 111914720 A CN111914720 A CN 111914720A CN 202010732376 A CN202010732376 A CN 202010732376A CN 111914720 A CN111914720 A CN 111914720A
Authority
CN
China
Prior art keywords
insulator
burst
transmission line
power transmission
sample
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202010732376.7A
Other languages
Chinese (zh)
Other versions
CN111914720B (en
Inventor
詹炜
朱晨光
舒子桓
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Yangtze University
Original Assignee
Yangtze University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Yangtze University filed Critical Yangtze University
Priority to CN202010732376.7A priority Critical patent/CN111914720B/en
Publication of CN111914720A publication Critical patent/CN111914720A/en
Application granted granted Critical
Publication of CN111914720B publication Critical patent/CN111914720B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/10—Terrestrial scenes
    • G06V20/13—Satellite images
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/20—Image preprocessing
    • G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
    • G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds

Abstract

The invention relates to the technical field of power inspection, and discloses a method for identifying insulator burst of a power transmission line, which comprises the following steps: acquiring an image of the power transmission line, and marking an insulator region in the image of the power transmission line to obtain an insulator marking frame; segmenting the image of the power transmission line to obtain a plurality of anchor frames, selecting a part of anchor frames, classifying the selected anchor frames according to the insulator marking frames to obtain a positive sample including the insulator and a negative sample not including the insulator; training the area proposed network by using the positive sample and the negative sample to obtain an insulator identification model; classifying all anchor frames according to the insulator identification model to obtain a positive sample and a negative sample, and marking a bursting area of the positive sample to obtain a training sample; training the neural network by using the training sample to obtain a burst recognition model; and carrying out burst recognition on the image to be recognized by combining the insulator recognition model and the burst recognition model. The invention has the technical effect of high polling efficiency.

Description

Method and device for identifying insulator burst of power transmission line
Technical Field
The invention relates to the technical field of power inspection, in particular to a method and a device for identifying insulator burst of a power transmission line and a computer storage medium.
Background
With the development of science and technology and the progress of society, the national demand for electric power is greater and greater, and the requirement for power supply reliability is higher and higher. At present, the scale of a power grid in China is the first in the world, and six power grid systems across provinces are in total, so that the large scale and the complex geographic environment of the power grid systems present a severe challenge to the reliability of power supply. The electric power inspection by only manual mode can not completely meet the extensive demand of modern electric power system, so the research strength of novel electric power inspection technology is increased at home and abroad and breakthrough progress is obtained, if: manned helicopter patrols and examines, and the robot patrols and examines, and the fixed wing unmanned aerial vehicle patrols and examines, and many rotor unmanned aerial vehicles patrol and examine novel electric power inspection technique such as waiting. Wherein multi-rotor unmanned aerial vehicles are most widely used. In 2017, each power grid company in China is provided with more than 2000 unmanned aerial vehicles, the total number of towers which are inspected by the unmanned aerial vehicles exceeds 20 million, the problems existing in inspection exceed 4 million, and the inspection work has a non-trivial role, and the basic work flow is as follows:
the working area of the unmanned aerial vehicle is defined, and the patrol time is set;
the unmanned aerial vehicle flies to a designated area and takes pictures (360 degrees of omnibearing is required to be achieved as far as possible);
analyzing pictures shot by the unmanned aerial vehicle, marking problems, and feeding back the problems to a dispatching center;
and the dispatching center arranges corresponding personnel for overhauling according to the problems.
However, because the number of pictures taken by the unmanned aerial vehicle is large (the number of pictures taken by a single overhead tower is larger than 300), the size is large (4096 × 2160 for example), 5-10 minutes are required for manually marking one picture, and the workload is huge. Meanwhile, related personnel who execute the labeling work are easy to use eyestrain, so that label leakage and label error are caused.
Disclosure of Invention
The invention aims to overcome the technical defects, provides a method and a device for identifying the insulator burst of the power transmission line and a computer storage medium, and solves the technical problems that in the prior art, the efficiency is low, the workload is large, and the marking error is easy to occur when the circuit inspection photos are identified and marked manually.
In order to achieve the technical purpose, the technical scheme of the invention provides a method for identifying the insulator burst of the power transmission line, which comprises the following steps:
acquiring an image of the power transmission line, and marking an insulator region in the image of the power transmission line to obtain an insulator marking frame;
segmenting the power transmission line image to obtain a plurality of anchor frames, selecting a part of anchor frames, classifying the selected anchor frames according to the insulator marking frames to obtain a positive sample including the insulator and a negative sample not including the insulator;
training the area proposed network by using the positive sample and the negative sample to obtain an insulator identification model;
classifying all the anchor frames according to the insulator recognition model to obtain a positive sample and a negative sample, and marking a burst area of the positive sample to obtain a training sample;
training the neural network by using the training sample to obtain a burst recognition model;
and carrying out burst identification on the image to be identified by combining the insulator identification model and the burst identification model.
The invention also provides a device for identifying the insulator burst of the power transmission line, which comprises a processor and a memory, wherein the memory is stored with a computer program, and the computer program is executed by the processor to realize the method for identifying the insulator burst of the power transmission line.
The invention also provides a computer storage medium, wherein a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the method for identifying the insulator burst of the power transmission line is realized.
Compared with the prior art, the invention has the beneficial effects that: the method mainly realizes the detection of the damage of the outer contour of the insulator by a deep learning method. Firstly, preliminarily segmenting the acquired electric transmission line image. And then extracting the insulator string by using a neural network. Finally, the neural network is reused to identify whether the extracted insulator string has an outline damage or not, the accuracy of marking the self-explosion insulator can be better improved, manual assistance is not needed, the probability of label leakage and label error is greatly reduced, and the power inspection work efficiency is improved.
Drawings
Fig. 1 is a flowchart of an embodiment of a method for identifying a burst insulator of a power transmission line according to the present invention;
FIG. 2 is a graph illustrating the results of one embodiment of a first predicted result provided by the present invention;
FIG. 3 is a diagram illustrating results of one embodiment of a second predicted result provided by the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
Example 1
As shown in fig. 1, embodiment 1 of the present invention provides a method for identifying a burst of an insulator of a power transmission line, including the following steps:
s1, collecting the image of the power transmission line, and marking the insulator region in the image of the power transmission line to obtain an insulator marking frame;
s2, segmenting the power transmission line image to obtain a plurality of anchor frames, selecting a part of anchor frames, and classifying the selected anchor frames according to the insulator marking frames to obtain a positive sample including the insulator and a negative sample not including the insulator;
s3, training the area proposal network by using the positive sample and the negative sample to obtain an insulator identification model;
s4, classifying all the anchor frames according to the insulator recognition model to obtain a positive sample and a negative sample, and labeling burst areas of the positive sample to obtain a training sample;
s5, training the neural network by using the training sample to obtain a burst recognition model;
and S6, performing burst recognition on the image to be recognized by combining the insulator recognition model and the burst recognition model.
In the embodiment, the images of the power transmission line are collected firstly, and in order to make the training model more accurate, the collected original images are expanded by adopting a method of turning over, changing the scale and changing the contrast to generate similar image data so as to obtain a power transmission line image library. Data annotation: and marking the power transmission line image, wherein the power transmission line image comprises insulators detected in the first round and insulator burst areas detected in the second round. And after marking is finished, constructing a network, training, and firstly training an insulator recognition model through a part of anchor frames to realize automatic recognition and classification of mass anchor frames. And training a burst recognition model according to the label of the burst area. And combining the two models to realize burst identification. The training uses batch processing, and one batch can process multiple pictures at a time, so the sizes of the pictures are required to be consistent. If the size of the input picture does not meet the required size, the network will scale the longer edge to the set pixel and fill the shorter edge with 0 to the set pixel.
The embodiment realizes the detection of the damage of the outer contour of the insulator through a deep learning method, can better improve the accuracy of identifying the self-explosion insulator, does not need manual assistance, greatly reduces the probability of label leakage and label error, and improves the power inspection work efficiency.
Preferably, the power transmission line image is segmented to obtain a plurality of anchor frames, and specifically, the method comprises the following steps:
inputting the power transmission line image into a residual error network to obtain a multi-dimensional characteristic diagram;
the high-level feature map is subjected to up-sampling and then is added and fused with the corresponding low-level feature map according to elements to obtain a fused feature map;
and carrying out pixel-by-pixel sliding window scanning on the feature map to generate a plurality of anchor frames.
In the embodiment, the backbone network adopts the residual error network, so that the problem of performance degradation caused by network deepening is effectively solved. The basic unit of the residual network establishes a direct path of correlation between input and output, a structure also referred to as a 'transition'. If the number of input and output channels is different, the residual network uses convolution of 1x1 to keep the input channels consistent. If the size of the input picture does not meet the required size, the residual network keeps the number of longer channels consistent with the number of output channels. After the pictures enter a backbone network, a multi-dimensional feature map is output at each STAGE, in order to have the advantages of strong semantic meaning of the high-level feature map, accurate position information of the low-level feature map and high resolution ratio of detectable small objects, a feature pyramid network structure is adopted, and the network structure is formed by adding and fusing the high-level feature map, which is subjected to upsampling and convolution with 1X1, with the same number of channels according to elements. Specifically, in this embodiment, the high-low level feature maps are fused by an upsampling and fusion method, 3 fused scales are 1313, 2626 and 5252, respectively, and the fusion method is that a small-scale image is doubled and added with a large-scale image element, and the detection is performed on the multiple-scale fused feature maps independently. It should be understood that the upper and lower layers are relative concepts herein. And after the characteristic graphs are obtained through fusion, scanning through a sliding window to generate an anchor frame.
Preferably, the feature map is scanned by a sliding window pixel by pixel to generate a plurality of anchor frames, specifically:
setting various length-width ratios of the anchor frame;
setting corresponding anchor frame sizes for each layer of characteristic diagram, and combining various length-width ratios to obtain various anchor frame sizes corresponding to each layer of characteristic diagram;
and traversing each layer of feature map, and performing pixel-by-pixel sliding window scanning on each layer of feature map to generate anchor frame groups with various sizes.
And traversing each layer of feature map, and performing pixel-by-pixel sliding window scanning on each layer of feature map to generate an anchor frame, wherein each layer of feature map corresponds to one size of anchor frame, and each size of anchor frame corresponds to three different length-width ratios of [0.5, 1 and 2 ]. Taking an original image of 1024 × 1024 size as an example, a total of about 26 ten thousand anchor frames will be generated.
Preferably, the selected anchor frame is classified according to the insulator marking frame to obtain a positive sample including the insulator and a negative sample not including the insulator, and the method specifically comprises the following steps:
and taking the insulator marking frame as a real frame, calculating the intersection ratio of the real frame and each anchor frame, dividing the anchor frame with the intersection ratio larger than a first set value into positive samples, calculating the offset between the positive samples and the corresponding real frame, and dividing the anchor frame with the intersection ratio smaller than a second set value into negative samples.
After 26 ten thousand anchor frames are obtained, a training area is required to propose network learning to distinguish which anchor frames contain insulators and which do not contain insulators, so that automatic classification of massive anchor frames is realized. An anchor box in which the ratio of the intersection area of the anchor box and the real box to the sum of the two areas (i.e., IoU intersection ratio) is greater than 0.7 is called a positive sample, IoU is less than 0.3 is called a negative sample, and the rest is a useless sample. Then 128 samples are selected from the positive samples and 128 samples are selected from the negative samples, and 256 samples are used for training the mark anchor box. Since there is also some deviation between the anchor frame marked as a positive sample and the real box, in order to make the prediction box closer to the real box, it is also necessary to learn the offset between the anchor frame and the real box from these samples.
Calculating the offset between the positive sample and the corresponding real frame, specifically as follows:
Gx＝ΔX+Px
Gy＝ΔY+Py
Gw＝Pwexp(dw(P))
Gh＝Phexp(dh(P))
ΔX＝Pwdx(P)
ΔY＝Phdy(P)
wherein G isxX value, G, being the coordinate of the center point of the real frameyY value, G, being the coordinates of the center point of the real framewIs the width of the real frame, GhIs the height of the real frame, PwWidth of positive sample, exp (d)w(P)) is the width scaling, PxX value, P, being the centre point coordinate of the positive sampleyIs the value of Y at the center point coordinate of the positive sample, deltax is the offset on the X-axis,Δ Y is the offset on the Y axis;
the following formulas are combined to obtain:
Figure BDA0002603657980000061
wherein phi is5Proposing a feature vector of the network for the input area, W*In order for the parameters to be learned,
Figure BDA0002603657980000062
is W*Comprises four parts (x, y, w, h), d*(P) denotes a prediction block. The purpose of the training is to let d*Offset t between (P) and real frame*Minimum difference, t*Comprises (t)x,ty,tw,th) I.e. loss value calculated by the loss function is minimal:
Figure BDA0002603657980000063
calculating the loss value by forward propagation, updating the weight by backward propagation
Figure BDA0002603657980000064
To achieve the purpose of training parameters, i is the serial number of the sample, and N is the number of samples participating in training, which is 256 in this embodiment.
Preferably, the bursting area of the positive sample is labeled to obtain a training sample, specifically:
selecting a set number of positive samples with the highest confidence rate as a target area, and adding the offset of the positive samples to the coordinates of the target area;
carrying out maximum suppression operation on the target area, and screening out repeated target areas corresponding to the same target;
if the number of the residual target areas is less than the set number after the repeated target areas are screened out, supplementing the target areas in a zero supplementing mode;
further screening the target area, removing the zero-filled target area, removing the target area containing a plurality of objects to be detected, and selecting a set number of positive samples containing the burst insulators and negative samples not containing the burst insulators from the rest target areas according to a set proportion;
and assigning the class label of the burst labeling frame with the maximum cross-over ratio with the positive sample to the positive sample, and calculating the offset between the positive sample and the real frame to obtain the training sample.
The obtained 26 ten thousand anchor frames trained area proposal network can be divided into positive samples and negative samples, 2000 positive samples with the highest score (confidence rate) are selected as an area of interest (ROI), and the offset is correspondingly added to the coordinates of the selected 2000 anchor frames, so that the ROI area is more accurate
The 2000 ROIs are subjected to a maximum suppression operation (NMS) before returning, and since one target may correspond to multiple ROIs, we need to remove the redundant ROIs. The ROI with the highest confidence rate is selected from the 2000 ROIs, and the intersection ratio between the other ROIs and the ROI with the highest confidence rate is calculated (IoU), and if IoU is greater than 0.7, the ROIs are excluded, and the above operations are repeated. After NMS operation, if ROI is less than 2000, it is filled in by zero padding.
In order to make the prediction box more accurate, a second fine tuning is required. Firstly, further screening 2000 anchor frames after first fine tuning and NMS operation, and removing ROIs filled with zeros; removing the ROI bounding multiple objects in the real box, and in the remaining ROI, if IoU of the ROI and the real box is greater than 0.5, setting as positive samples, IoU is less than 0.5, setting as negative samples, wherein the positive samples are set to 50, the negative samples are set to 150, and the positive and negative samples keep 1: 3, in the presence of a catalyst.
For 50 positive samples, the real box with the maximum of IoU is calculated, and the class label of the real box is assigned to the positive sample; and calculating the offset between the positive samples and the real box as in the process of training the area proposal network, and returning the category and the offset of the positive samples.
Preferably, the maximum suppression operation is performed on the target area to screen out a repeated target area corresponding to the same target, specifically:
and calculating the intersection ratio between the target area with the highest confidence rate and other target areas, and removing the corresponding target area if the intersection ratio is greater than a set threshold.
Preferably, the labeling of the burst region of the positive sample is performed to obtain a training sample, and the method further includes: carrying out secondary adjustment on the training sample;
carrying out secondary adjustment on the training sample, specifically comprising the following steps:
searching a characteristic diagram corresponding to the training sample, and calculating the size of the training sample on the corresponding characteristic diagram;
dividing a region corresponding to a training sample on the feature map into a plurality of small regions, dividing each small region into four equal parts, calculating central point pixels of the four equal parts in each small region by adopting a bilinear interpolation method, and taking the maximum value of the four calculated central point pixels as the pixel value of the corresponding small region;
combining the pixel values of the small regions to obtain a feature block corresponding to the training sample;
and inputting the feature block into a full-link layer and a softmax layer for classification, and returning to a corresponding class label and a real frame coordinate of the training sample.
The first ROI fine adjustment is performed, the ROI is mapped to the feature map from the original image, and then the ROI is mapped to the original image from the feature map, so that stride errors are generated in the process. To reduce this error, the second positive sample fine-tuning uses the ROI alignment method. The method comprises the steps of firstly mapping a positive sample to a corresponding feature map, dividing the length and width of the positive sample by the step length of the feature map to obtain the size of the positive sample on the feature map, then dividing the positive sample area on the feature map into 49 small areas of 7x7, dividing each small area into 4 equal parts, then calculating the central point pixel of the 4 equal parts by a bilinear interpolation method, taking the maximum value of 4 pixels as the pixel value of the small area, sequentially taking 49 pixel values of the 49 small areas, and forming the feature map of 7x7 by the 49 pixels. In the ROI alignment method, quantification is not performed in the operation process, floating point calculation is adopted to reserve decimal, and the error of mapping the feature graph to the original graph is smaller.
Searching a characteristic diagram corresponding to the positive sample, specifically:
Figure BDA0002603657980000081
wherein w represents the width of the positive sample, h represents the height of the positive sample, K is the number of the feature map layers corresponding to the positive sample, and K0The number of feature map layers corresponding to (w, h) being (224 );
calculate the size of the positive sample on the corresponding feature map:
and dividing the width and the height of the positive sample by the step size of the corresponding characteristic diagram respectively to obtain the size of the positive sample on the characteristic diagram.
The value (224 ) of (w, h) in this embodiment is because the standard picture size of ImageNet is 224x224, for example, if there is an ROI of 112 × 112, k may be calculated to be 3, that is, P3 layers by using the formula.
The 7 × 7 feature map is obtained and sorted via full connectivity layer and softmax back to class _ id and real box coordinates.
Preferably, the to-be-recognized image is subjected to burst recognition by combining the insulator recognition model and the burst recognition model, and the method specifically comprises the following steps:
the method comprises the steps of reducing an image to be detected in proportion, inputting the reduced image to be detected into a burst identification model, and obtaining a first prediction result;
preprocessing the image to be detected, inputting the preprocessed image to be detected into the insulator identification model to obtain an insulator region, and judging whether the insulator region has insulator burst or not to obtain a second prediction result;
judging whether the first prediction result is coincident with the second prediction result, if so, outputting the first prediction result or the second prediction result, if not, expanding the first prediction result to obtain an expansion matrix, further judging whether the expansion matrix is multiple, if so, outputting the first prediction result, and if not, outputting the second prediction result.
After the model training is finished, automatic identification can be carried out. Because the image definition is high and the neural network learning process is too long, the image is firstly scaled down in the process to shorten the learning process, and the scaled-down image to be detected is input into the burst recognition model to obtain a first prediction result of the burst region of the insulator, as shown in fig. 2, a black frame in the image is the first prediction result.
And preprocessing the image to be recognized so as to facilitate the deep learning process of the two models on the image. Inputting the preprocessed image to be detected into an insulator recognition model to obtain an insulator region, and then inputting the insulator region into a burst recognition model, wherein the process is actually to narrow the range of a candidate region of the spontaneous explosion region, intercept the insulator region obtained in the insulator recognition model, judge whether the insulator is damaged in the region, and obtain a second prediction result of the insulator spontaneous explosion region, as shown in fig. 3, a black frame in the figure is the second prediction result.
Due to the mode problem of image labeling, a phenomenon that a plurality of insulators are regarded as one insulator exists in the first stage, when the positions in the prediction result (1) are not overlapped in the position (2), the extended matrix formed by the positions in the positions (1) needs to be judged again, if a plurality of matrixes exist, the prediction result (1) is adopted, and otherwise, the prediction result (2) is adopted.
The preferred embodiment realizes further improvement of the identification accuracy by utilizing the mutual correction of the prediction results of the two models.
Example 2
Embodiment 2 of the present invention provides a device for identifying a burst of an insulator of a power transmission line, which includes a processor and a memory, where the memory stores a computer program, and when the computer program is executed by the processor, the method for identifying a burst of an insulator of a power transmission line provided in embodiment 1 is implemented.
The device for identifying the insulator burst of the power transmission line provided by the embodiment of the invention is used for realizing the method for identifying the insulator burst of the power transmission line, so that the device for identifying the insulator burst of the power transmission line has the technical effect, and the device for identifying the insulator burst of the power transmission line also has the technical effect, and is not repeated herein.
Example 3
Embodiment 3 of the present invention provides a computer storage medium having a computer program stored thereon, where the computer program, when executed by a processor, implements the method for identifying a burst insulator of a power transmission line provided in embodiment 1.
The computer storage medium provided by the embodiment of the invention is used for realizing the method for identifying the insulator burst of the power transmission line, so that the computer storage medium has the technical effects of the method for identifying the insulator burst of the power transmission line, and the description is omitted.
The above-described embodiments of the present invention should not be construed as limiting the scope of the present invention. Any other corresponding changes and modifications made according to the technical idea of the present invention should be included in the protection scope of the claims of the present invention.

Claims (10)

1. A method for identifying insulator burst of a power transmission line is characterized by comprising the following steps:
acquiring an image of the power transmission line, and marking an insulator region in the image of the power transmission line to obtain an insulator marking frame;
segmenting the power transmission line image to obtain a plurality of anchor frames, selecting a part of anchor frames, classifying the selected anchor frames according to the insulator marking frames to obtain a positive sample including the insulator and a negative sample not including the insulator;
training the area proposed network by using the positive sample and the negative sample to obtain an insulator identification model;
classifying all the anchor frames according to the insulator recognition model to obtain a positive sample and a negative sample, and marking a burst area of the positive sample to obtain a training sample;
training the neural network by using the training sample to obtain a burst recognition model;
and carrying out burst identification on the image to be identified by combining the insulator identification model and the burst identification model.
2. The method for identifying the insulator burst of the power transmission line according to claim 1, wherein the image of the power transmission line is segmented to obtain a plurality of anchor frames, and specifically comprises the following steps:
inputting the power transmission line image into a residual error network to obtain a multi-dimensional characteristic diagram;
the high-level feature map is subjected to up-sampling and then is added and fused with the corresponding low-level feature map according to elements to obtain a fused feature map;
and carrying out pixel-by-pixel sliding window scanning on the feature map to generate a plurality of anchor frames.
3. The method for identifying the insulator burst of the power transmission line according to claim 1, wherein the characteristic diagram is subjected to sliding window scanning pixel by pixel to generate a plurality of anchor frames, specifically:
setting various length-width ratios of the anchor frame;
setting corresponding anchor frame sizes for each layer of characteristic diagram, and combining various length-width ratios to obtain various anchor frame sizes corresponding to each layer of characteristic diagram;
and traversing each layer of feature map, and performing pixel-by-pixel sliding window scanning on each layer of feature map to generate anchor frame groups with various sizes.
4. The method for identifying the burst insulator of the power transmission line according to claim 1, wherein the selected anchor frame is classified according to the insulator marking frame to obtain a positive sample including the insulator and a negative sample not including the insulator, and specifically comprises the following steps:
and taking the insulator marking frame as a real frame, calculating the intersection ratio of the real frame and each anchor frame, dividing the anchor frame with the intersection ratio larger than a first set value into positive samples, calculating the offset between the positive samples and the corresponding real frame, and dividing the anchor frame with the intersection ratio smaller than a second set value into negative samples.
5. The method for identifying the electric transmission line insulator burst according to claim 1, wherein the burst area of the positive sample is labeled to obtain a training sample, and specifically comprises the following steps:
selecting a set number of positive samples with the highest confidence rate as a target area, and adding the offset of the positive samples to the coordinates of the target area;
carrying out maximum suppression operation on the target area, and screening out repeated target areas corresponding to the same target;
if the number of the residual target areas is less than the set number after the repeated target areas are screened out, supplementing the target areas in a zero supplementing mode;
further screening the target area, removing the zero-filled target area, removing the target area containing a plurality of objects to be detected, and selecting a set number of positive samples containing the burst insulators and negative samples not containing the burst insulators from the rest target areas according to a set proportion;
and assigning the class label of the burst labeling frame with the maximum cross-over ratio with the positive sample to the positive sample, and calculating the offset between the positive sample and the real frame to obtain the training sample.
6. The method for identifying the insulator burst of the power transmission line according to claim 5, wherein the maximum suppression operation is performed on the target area to screen out repeated target areas corresponding to the same target, specifically:
and calculating the intersection ratio between the target area with the highest confidence rate and other target areas, and removing the corresponding target area if the intersection ratio is greater than a set threshold.
7. The method for identifying the electric transmission line insulator burst according to claim 1, wherein the burst area of the positive sample is labeled to obtain a training sample, and the method further comprises the following steps: carrying out secondary adjustment on the training sample;
carrying out secondary adjustment on the training sample, specifically comprising the following steps:
searching a characteristic diagram corresponding to the training sample, and calculating the size of the training sample on the corresponding characteristic diagram;
dividing a region corresponding to a training sample on the feature map into a plurality of small regions, dividing each small region into four equal parts, calculating central point pixels of the four equal parts in each small region by adopting a bilinear interpolation method, and taking the maximum value of the four calculated central point pixels as the pixel value of the corresponding small region;
combining the pixel values of the small regions to obtain a feature block corresponding to the training sample;
and inputting the feature block into a full-link layer and a softmax layer for classification, and returning to a corresponding class label and a real frame coordinate of the training sample.
8. The method for identifying the burst of the insulator of the power transmission line according to claim 1, wherein the burst identification is performed on the image to be identified by combining the insulator identification model and the burst identification model, and specifically comprises the following steps:
the method comprises the steps of reducing an image to be detected in proportion, inputting the reduced image to be detected into a burst identification model, and obtaining a first prediction result;
preprocessing the image to be detected, inputting the preprocessed image to be detected into the insulator identification model to obtain an insulator region, and judging whether the insulator region has insulator burst or not to obtain a second prediction result;
judging whether the first prediction result is coincident with the second prediction result, if so, outputting the first prediction result or the second prediction result, if not, expanding the first prediction result to obtain an expansion matrix, further judging whether the expansion matrix is multiple, if so, outputting the first prediction result, and if not, outputting the second prediction result.
9. An insulator burst recognition device for a power transmission line, which is characterized by comprising a processor and a memory, wherein the memory stores a computer program, and the computer program is executed by the processor to realize the insulator burst recognition method for the power transmission line according to any one of claims 1-8.
10. A computer storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the method for identifying a power transmission line insulator burst according to any one of claims 1 to 8.
CN202010732376.7A 2020-07-27 2020-07-27 Method and device for identifying insulator burst of power transmission line Active CN111914720B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010732376.7A CN111914720B (en) 2020-07-27 2020-07-27 Method and device for identifying insulator burst of power transmission line

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010732376.7A CN111914720B (en) 2020-07-27 2020-07-27 Method and device for identifying insulator burst of power transmission line

Publications (2)

Publication Number Publication Date
CN111914720A true CN111914720A (en) 2020-11-10
CN111914720B CN111914720B (en) 2021-08-17

Family

ID=73280862

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010732376.7A Active CN111914720B (en) 2020-07-27 2020-07-27 Method and device for identifying insulator burst of power transmission line

Country Status (1)

Country Link
CN (1) CN111914720B (en)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112464756A (en) * 2020-11-13 2021-03-09 上海电力大学 Image quantification method for insulator defect identification
CN113256604A (en) * 2021-06-15 2021-08-13 广东电网有限责任公司湛江供电局 Insulator string defect identification method and equipment for double learning
CN113420648A (en) * 2021-06-22 2021-09-21 深圳市华汉伟业科技有限公司 Target detection method and system with rotation adaptability
CN114841993A (en) * 2022-05-31 2022-08-02 广东电网有限责任公司 Training of insulator detection network, detection method and equipment thereof, and storage medium
CN117095011A (en) * 2023-10-20 2023-11-21 南通华隆微电子股份有限公司 Diode detection method and system

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106127756A (en) * 2016-06-21 2016-11-16 西安工程大学 A kind of insulator recognition detection method based on multicharacteristic information integration technology
CN107274451A (en) * 2017-05-17 2017-10-20 北京工业大学 Isolator detecting method and device based on shared convolutional neural networks
CN108280855A (en) * 2018-01-13 2018-07-13 福州大学 A kind of insulator breakdown detection method based on Fast R-CNN

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106127756A (en) * 2016-06-21 2016-11-16 西安工程大学 A kind of insulator recognition detection method based on multicharacteristic information integration technology
CN107274451A (en) * 2017-05-17 2017-10-20 北京工业大学 Isolator detecting method and device based on shared convolutional neural networks
CN108280855A (en) * 2018-01-13 2018-07-13 福州大学 A kind of insulator breakdown detection method based on Fast R-CNN

Cited By (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112464756A (en) * 2020-11-13 2021-03-09 上海电力大学 Image quantification method for insulator defect identification
CN112464756B (en) * 2020-11-13 2023-05-02 上海电力大学 Insulator defect identification-oriented image quantization method
CN113256604A (en) * 2021-06-15 2021-08-13 广东电网有限责任公司湛江供电局 Insulator string defect identification method and equipment for double learning
CN113420648A (en) * 2021-06-22 2021-09-21 深圳市华汉伟业科技有限公司 Target detection method and system with rotation adaptability
CN114841993A (en) * 2022-05-31 2022-08-02 广东电网有限责任公司 Training of insulator detection network, detection method and equipment thereof, and storage medium
CN117095011A (en) * 2023-10-20 2023-11-21 南通华隆微电子股份有限公司 Diode detection method and system
CN117095011B (en) * 2023-10-20 2024-01-23 南通华隆微电子股份有限公司 Diode detection method and system

Also Published As

Publication number Publication date
CN111914720B (en) 2021-08-17

Similar Documents

Publication Publication Date Title
CN111914720B (en) Method and device for identifying insulator burst of power transmission line
CN108961235B (en) Defective insulator identification method based on YOLOv3 network and particle filter algorithm
CN109685067B (en) Image semantic segmentation method based on region and depth residual error network
CN110728200B (en) Real-time pedestrian detection method and system based on deep learning
CN110059694B (en) Intelligent identification method for character data in complex scene of power industry
CN109118479B (en) Capsule network-based insulator defect identification and positioning device and method
CN112199993B (en) Method for identifying transformer substation insulator infrared image detection model in any direction based on artificial intelligence
CN111914795B (en) Method for detecting rotating target in aerial image
CN111784685A (en) Power transmission line defect image identification method based on cloud edge cooperative detection
Wang et al. Object-scale adaptive convolutional neural networks for high-spatial resolution remote sensing image classification
CN113378686B (en) Two-stage remote sensing target detection method based on target center point estimation
CN111967313B (en) Unmanned aerial vehicle image annotation method assisted by deep learning target detection algorithm
CN110781882A (en) License plate positioning and identifying method based on YOLO model
CN115205264A (en) High-resolution remote sensing ship detection method based on improved YOLOv4
CN112560675A (en) Bird visual target detection method combining YOLO and rotation-fusion strategy
CN111768415A (en) Image instance segmentation method without quantization pooling
CN116385958A (en) Edge intelligent detection method for power grid inspection and monitoring
CN115272204A (en) Bearing surface scratch detection method based on machine vision
CN115861619A (en) Airborne LiDAR (light detection and ranging) urban point cloud semantic segmentation method and system of recursive residual double-attention kernel point convolution network
CN112785610B (en) Lane line semantic segmentation method integrating low-level features
CN114529821A (en) Offshore wind power safety monitoring and early warning method based on machine vision
CN114445615A (en) Rotary insulator target detection method based on scale invariant feature pyramid structure
CN113326734A (en) Rotary target detection method based on YOLOv5
CN109871903B (en) Target detection method based on end-to-end deep network and counterstudy
CN116342496A (en) Abnormal object detection method and system for intelligent inspection

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant