A method for detecting self-explosion of insulator string of power transmission line based on deep background extinction

Abstract

The application discloses a kind of based on depth background extinction transmission line insulator string self-explosion detection method, comprising the following steps: S1: by threshold method to aerial image is handled, obtains the binary mask of insulator string;S2: by image erosion operation removes noise, determine the main part of insulator string, again by image inflation to the main part of insulator string is divided into uncertain area, obtains Trimap chart;S3: Trimap chart is input into depth neural network and inquires uncertain area, obtains Alpha chart;S4: using Alpha chart will original drawing from RGB three channels to RGBA four channels, obtains after background extinction insulator string image;S5: with light YOLOv5 algorithm is carried out to the detection of insulator string self-explosion defect.The application can remove the complex background where power equipment is located, greatly reduce the adverse effects caused by transmission line complex environment on detection precision, meet the detection requirements of insulator string self-explosion defect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of self-explosion detection of power transmission line insulator strings, and particularly relates to a self-explosion detection method for power transmission line insulator strings based on deep background extinction. BACKGROUND

[0002] In recent years, with the continuous expansion of the scale of power transmission lines in China, the health problems of power transmission line equipment cannot be ignored. As one of the most common power equipment in high-voltage power transmission lines, the insulator string plays an important role in electrical insulation and mechanical support. However, the insulator string of the power transmission line is exposed to the field for a long time during operation, and is prone to damage, self-explosion and other faults due to natural conditions such as rain corrosion and lightning, which seriously affects the safe and stable operation of the power grid.

[0003] Traditional insulator string defect detection is generally performed by on-site inspection by inspection personnel. However, due to the distribution range and environment of the power transmission line, manual inspection is inefficient. Under the wave of digitalization and intelligentization of the power industry, unmanned aerial vehicle inspection has become the future development trend of the power inspection industry, and high-precision detection of self-explosion faults of the insulator string by the computer is a key prerequisite for unmanned aerial vehicle inspection.

[0004] The existing power transmission line defect detection methods mainly include the Two-stage (two-stage) algorithm represented by Faster-RCNN and the One-stage (single-stage) algorithm represented by YOLO (You Only Look Once). The RPN (region proposal network) proposed by the Faster-RCNN algorithm performs a preliminary detection on the image first, which can better find the target region. However, the presence of RPN in Faster-RCNN greatly reduces the detection speed, which cannot meet the requirement of high efficiency of power inspection. Therefore, the current inspection unmanned aerial vehicle is mostly equipped with the One-stage YOLO series algorithm. Although YOLO improves the detection speed by dividing the detection image into a grid, it has the advantage of real-time performance, but its disadvantage is also very significant: since the YOLO series algorithm is a single-stage detection, the speed is improved at the expense of part of the accuracy. Although it can still meet the precision requirement of fault detection in a simple background, for power transmission line defect images in a complex environment, the defect detection precision of the YOLO algorithm is unsatisfactory. SUMMARY

[0005] The technical problem to be solved by the present application is to provide a self-explosion detection method for power transmission line insulator strings based on deep background extinction, which can realize accurate detection of insulator string self-explosion defects in complex environments.

[0006] To solve the above technical problems, the technical solutions adopted by the present application are as follows.

[0007] A power transmission line insulator string self-explosion detection method based on deep background extinction, comprising the following steps:

[0008] S1: The threshold method is used to process the aerial image to obtain a binary mask image of the insulator string;

[0009] S2: The noise is removed by image erosion operation to determine the main part of the insulator string, and the main part of the insulator string is divided into an uncertain area by image dilation to obtain a Trimap image;

[0010] S3: The Trimap image is input into a deep neural network for inference of the uncertain area to obtain an Alpha image;

[0011] S4: The original image is expanded from RGB three channels to RGBA four channels using the Alpha image to obtain an insulator string image after background extinction;

[0012] S5: A lightweight YOLOv5 algorithm is used to detect the insulator string self-explosion defect.

[0013] Preferably, the image erosion is realized by the function cv:erode() of OpenCV, and the image dilation is realized by the function cv:dilate() of OpenCV in step S2.

[0014] Preferably, the specific method for obtaining the Alpha image in step S3 is:

[0015] The original image is denoted as RGB, the Trimap image is denoted as Trimap, and the splicing of the original image and the Trimap image is denoted as RGB-Trimap, which are respectively input into different neural networks for feature extraction;

[0016] For the RGB image, three consecutive convolution operations are used: the convolution kernel size is 3, the number of channels is 16, 64 and 128 respectively, and the span of the three convolutions is set to 2;

[0017] For the Trimap image, the ResNet-18 network is used to extract the high-level global information of the Trimap, and only the first three modules, i.e., conv-1, res-2 and res-3, are used, and the kernel size and the number of channels of the ResNet-18 are not changed;

[0018] For the RGB-Trimap image, the U-Net network is used for feature extraction;

[0019] After the feature extraction is completed, the features from the RGB-Trimap branch and the RGB and Trimap are input into the encoder-decoder for integration of the three-way information to obtain the Alpha image.

[0020] Preferably, the RGB-Trimap branch in the step S3 introduces ASPP to extract high-level semantic information to improve the expression ability of the RGB-Trimap branch.

[0021] Preferably, the specific method of the step S4 is:

[0022] S41: Channel separation is performed on the Alpha map through the function cv: split() of OpenCV, and pixel values of any one channel are extracted;

[0023] S42: The pixel values extracted in the step S41 are combined with the RGB three channels of the original image through the function cv: merge() of OpenCV, and an insulator string image after background extinction is obtained.

[0024] Thanks to the above technical solutions, the technical progress achieved by the present application is as follows.

[0025] The present application has high precision, and by adding a deep neural background extinction preprocessing network before the YOLOv5 target detection network, the complex background in which the power equipment is located can be removed, the adverse effects of the complex environment of the power transmission line on the detection precision are greatly reduced, and the detection requirements of the insulator string self-explosion defect are met.

[0026] The present application has strong generalization, and by using the method of artificial intelligence to detect the defects of the power transmission line equipment, the generalization ability of the model is significantly improved due to the existence of the background extinction preprocessing network. When facing new defects, the present application only needs to label the new defect aerial image and set it as an expanded data set, and the algorithm network can be migrated and trained, which greatly reduces the workload. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a flow chart of the present application;

[0028] Figure 2 is a deep neural background extinction preprocessing network structure diagram of the present application;

[0029] Figure 3 is a deep learning target detection network structure diagram of the present application;

[0030] Figure 4 is an insulator string mask image obtained by thresholding method of the present application;

[0031] Figure 5 is an insulator string Trimap image obtained after erosion and inflation processing of the present application;

[0032] Figure 6An insulator string Alpha image obtained after a deep neural background extinction preprocessing network of the application;

[0033] Figure 7 An insulator string image after background removal obtained after channel expansion of the application. DETAILED DESCRIPTION

[0034] The application will be further described in detail below in combination with the drawings and specific embodiments.

[0035] A power transmission line insulator string self-explosion detection method based on deep background extinction, which can realize accurate detection of insulator string self-explosion defects in complex environments by performing extinction preprocessing on complex backgrounds, combined with Figure 1 As shown in the drawings, the method comprises the following steps:

[0036] S1: Process the aerial image by threshold method to obtain a binary mask image (mask image) of the insulator string.

[0037] The mask image is a binary image in which the foreground of the image is marked as white and the background of the image is marked as black.

[0038] Since the colors of different types of insulator strings are monotonous, the overall color is uniform, and the color difference with the background is large, the threshold method can be used to obtain a rough binary mask image (mask image) of the insulator string.

[0039] We call the target to be detected (i.e. the insulator string) as foreground, and other objects and environment as background. Then for the foreground to be detected, for porcelain insulator, the color is generally deep red, in order to distinguish from the background, we set the threshold range of the foreground as: 200<R<255, 0<G<50, 0<B<50; for glass insulator, the color is generally sky blue, in order to distinguish from the background, we set the threshold range of the foreground as: 0<R<10, 150<G<225, 200<B<255; we identify the pixel points within the above threshold range as foreground, and set the pixel value as (255, 255, 255), i.e. white; identify the pixel points outside the above threshold range as background, and set the pixel value as (0, 0, 0), i.e. black, thereby obtaining the mask image as shown in Figure 4 .

[0040] S2: Remove noise by image erosion operation to determine the main part of the insulator string, and then divide the main part of the insulator string into uncertain regions by image dilation to obtain a Trimap image.

[0041] The Trimap is also called trimap, which is generated on the basis of the mask. In order to obtain a more fine binary mask, in addition to the foreground and background of the original mask, an uncertain area is added, which is marked by gray (127, 127, 127). The image divided into three parts of foreground, background and uncertain area is called trimap, that is, Trimap.

[0042] In order to obtain the Trimap, the key link is to obtain the uncertain area. The present application adopts the method of image erosion and image dilation to obtain the uncertain area.

[0043] Image erosion refers to the corrosion compression of the foreground part marked in the original mask. Since the original mask is obtained by threshold method, noise that does not belong to the target object but is within the target threshold range will inevitably appear, and the purpose of image erosion is to eliminate these noises and obtain the main part of the target object. The operation steps of image erosion are: convolve image A with an arbitrary shaped kernel (B), usually square or circular. Specifically, the kernel B has a definable anchor point, usually defined as the kernel center point, and the kernel B is drawn across the image A, and the minimum pixel value of the kernel B covering area is extracted and replaces the pixel at the anchor point position. The minimum value is obtained by the following formula:

[0044] dst(x,y) = min src(x+x′,y+y′)

[0045] Obviously, this minimization operation will cause the bright area in image A to start "shrinking". The present application realizes image erosion through the function cv: erode () of OpenCV.

[0046] Image dilation refers to the dilation of the pixel value of the main part after image erosion, that is, the division of the uncertain area. The operation steps of image dilation are similar to those of image erosion: convolve image A with an arbitrary shaped kernel (B), usually square or circular. Specifically, the kernel B has a definable anchor point, usually defined as the kernel center point, and the kernel B is drawn across the image A, and the maximum pixel value of the kernel B covering area is extracted and replaces the pixel at the anchor point position. The maximum value is obtained by the following formula:

[0047] dst(x,y) = max src(x+x′,y+y′)

[0048] Obviously, this minimization operation will cause the bright area in image A to start "shrinking". The present application realizes image erosion through the function cv: erode () of OpenCV.

[0049] Figure 5 is the Trimap obtained by eroding the original mask and dilating it by 30 pixel values.

[0050] S3: input the Trimap into the deep neural network to infer the uncertain area and obtain the Alpha map.

[0051] The Alpha map is an image used to indicate the opacity of the foreground and background of the image.

[0052] In order to obtain the Alpha map of the image, the present application uses a deep neural background extinction preprocessing network as shown in Figure 2

[0053] The original image (denoted as RGB), the Trimap (denoted as Trimap), and the splicing of the original image and the Trimap (denoted as RGB-Trimap) are respectively input into different neural networks for feature extraction.

[0054] For the RGB image, we use three consecutive convolution operations: the kernel size is 3, and the number of channels is 16, 64, and 128, respectively. In addition, considering the computational burden, we set the span of the three convolutions to 2, thereby achieving 8 times down-sampling, so as to balance the resolution and calculation. Through this operation, the global appearance can be preserved, especially the color and texture information from the RGB image, and when the foreground and background are locally similar, the ambiguity can be well eliminated.

[0055] For the Trimap image, we use a relatively deep network, ResNet-18, to extract the high-level global information of the Trimap. At the same time, we only use the first three modules (conv-1, res-2, res-3), and do not change the kernel size and the number of channels of ResNet-18.

[0056] For the RGB-Trimap image, we use the U-Net network for feature extraction, which is used as a baseline to extract the local information of the RGB-Trimap.

[0057] After the feature extraction is completed, the features from the RGB-Trimap branch and the other two units (RGB, Trimap) enter the encoder-decoder for the integration of the three-party information. At the same time, in order to improve the expression ability of the RGB-Trimap branch, we introduce the ASPP to extract high-level semantic information.

[0058] Figure 6 That is, the Alpha map output by the neural network.

[0059] ​S4: Use the Alpha image to expand the original image from RGB three channels to RGBA four channels to obtain the image of the insulator string after background extinction.

[0060] Background matte printing refers to an image where the background is removed, leaving only the foreground.

[0061] The original image has three channels: R, G, and B. To obtain a background extinction image, this invention expands the original image from RGB three channels to RGBA four channels, adding an Alpha transparency channel. This transparency channel originates from the Alpha image obtained in the third step: the Alpha image itself also has RGB three channels, and the pixel value of each channel represents the exact same opacity.

[0062] The specific method for obtaining the image of the insulator string after background extinction is as follows:

[0063] S41: Use OpenCV's cv::split() function to separate the channels of the alpha image and extract the pixel value of any one channel. The pixel value of that channel is the opacity we need.

[0064] S42: The pixel values ​​extracted in step S41 are merged with the RGB three channels of the original image using the OpenCV function cv::merge(), thus obtaining the image as shown below. Figure 7 The image shown is of the insulator string after removing the background, i.e., the background extinction image.

[0065] S5: Detection of insulator string self-explosion defects using the lightweight YOLOv5 algorithm.

[0066] like Figure 3 As shown, the target detection model used in this invention is the YOLOv5 algorithm. This algorithm uses CSPDarknet53 and FPN+PAN Net as the backbone network and neck part, respectively. In the prediction part, YOLOv5 includes three detection heads, used for the detection of small, medium, and large objects. When the background extinction image of the insulator string obtained in this invention is input into the YOLOv5 target detection model, the detection efficiency and accuracy of the insulator string are significantly improved due to the absence of interference from complex backgrounds.

[0067] In use, this invention adds a deep neural background extinction-based preprocessing network before the YOLOv5 target detection network, which can remove the complex background of the power equipment, greatly reducing the adverse effects of the complex environment of the transmission line on the detection accuracy and meeting the detection requirements of insulator string self-explosion defects.

[0068] The application adopts the method of artificial intelligence to detect defects of power equipment of a power transmission line. Due to the existence of the background extinction pretreatment network, the generalization ability of the model is significantly improved. When facing new defects, the application only needs to label new defect aerial images and set them as an expanded data set, and then migrates and trains the algorithm network, so that the workload is greatly reduced.

Claims

1. A method for detecting spontaneous explosion of transmission line insulator strings based on deep background extinction, characterized in that: Includes the following steps: S1: The aerial image is processed by thresholding to obtain a binary mask image of the insulator string; S2: Remove noise through image erosion to determine the main body of the insulator string, and then divide the uncertain regions of the main body of the insulator string through image dilation to obtain a Trimap image; S3: Input the Trimap graph into a deep neural network to perform inference on uncertain regions and obtain an Alpha graph; S4: Use the Alpha image to expand the original image from RGB three channels to RGBA four channels to obtain the image of the insulator string after background extinction; S41: Use the OpenCV function cv::split() to separate the channels of the alpha image and extract the pixel value of any channel; S42: The pixel values ​​extracted in step S41 are merged with the RGB three channels of the original image using the OpenCV function cv::merge() to obtain the insulator string image after background extinction. S5: Detection of insulator string self-explosion defects using the lightweight YOLOv5 algorithm.

2. The method for detecting spontaneous explosion of transmission line insulator strings based on deep background extinction as described in claim 1, characterized in that: In step S2, image erosion is achieved using the OpenCV function cv::erode(), and image dilation is achieved using the OpenCV function cv::dilate().

3. The method for detecting spontaneous explosion of transmission line insulator strings based on deep background extinction as described in claim 1, characterized in that: The specific method for obtaining the Alpha image in step S3 is as follows: The original image is denoted as RGB, the Trimap image as Trimap, and the concatenation of the original image and the Trimap image as RGB-Trimap. These are then input into different neural networks for feature extraction. For RGB images, three consecutive convolution operations are used: the kernel size is 3, the number of channels is 16, 64 and 128 respectively, and the span of the three convolutions is set to 2. For the Trimap graph, the ResNet-18 network is used to extract high-level global information of the Trimap, and only the first three modules, namely conv-1, res-2 and res-3, are used without changing the kernel size and number of channels of ResNet-18. For RGB-Trimap images, the U-Net network is used for feature extraction; After feature extraction is complete, the features from the RGB-Trimap branch and the RGB and Trimap branches are fed into the encoder-decoder for integration of the three-way information to obtain the Alpha image.

4. The method for detecting spontaneous explosion of transmission line insulator strings based on deep background extinction as described in claim 3, characterized in that: In step S3, ASPP is introduced into the RGB-Trimap branch to extract advanced semantic information and improve the expressive power of the RGB-Trimap branch.

Images