Horizontal rotating knife switch state accurate identification device and closing state judgment method

Abstract

The application discloses a horizontal rotary knife switch state accurate identification device and a closing state judgment method, which are used for identifying whether a horizontal rotary knife switch of a transformer substation is accurately closed in place. The device comprises a camera, a knife switch identification front-end device, a knife switch identification background learning and training device and a data and control bus. The data and control bus is used for transmitting data and control signals between the camera, the knife switch identification front-end device and the knife switch identification background learning and training device. The knife switch is an extremely important equipment in the transformer substation, and long-term use can easily lead to the knife switch not being closed in place, causing serious accidents such as heating and even explosion. The application adopts a cloud edge fusion architecture, uses a Mask R-CNN network to perform target identification, uses an HED method to perform edge extraction, and further proposes a method for detecting the knife switch not being closed in place, so that the accurate identification of the knife switch state is realized, manual inspection can be replaced, manpower can be saved, and safety can be improved.

Description

TECHNICAL FIELD

[0001] The application relates to power equipment detection technology, in particular to horizontal rotary knife switch state accurate detection, and more particularly to a horizontal rotary knife switch state accurate identification device and a closing state judgment method. BACKGROUND

[0002] The knife switch, also known as a high-voltage disconnector, is a main equipment of a substation. The knife switch described in the application is a single-phase knife switch. A three-phase knife switch has one knife switch for each phase, and the three knife switches are linked together. In the application, each knife switch has two horizontally rotatable arms, and the front ends of the arms have contacts.

[0003] In long-term operation, the knife switch may not be closed in place due to various reasons. The existing closing-in-place discrimination mainly relies on the auxiliary contact points of the knife switch. Since the knife switch is basically placed outdoors for work, it has to face the problems of sun exposure, rain and corrosion damage, and wear and tear. Abnormal auxiliary contact points or their transmission parts may cause misjudgment of the on-off position, upload of false signals, threat to equipment safety and service life, and even cause major power accidents, resulting in significant personnel casualties and property losses.

[0004] The traditional method for identifying the closing-in-place failure of the knife switch is to send inspection personnel to the scene to observe with the naked eye. However, since the substations are mostly distributed in remote areas, and the number of knife switches in each substation is large, on-site observation is time-consuming and labor-intensive. In addition, with the continuous development of society, the number of substations and knife switches used is also increasing. In order to quickly detect the closing conditions of a large number of knife switches at the same time, more human resources are needed compared to the past. This manual double detection method consumes a large amount of manpower and has poor real-time performance, which is a short board for the realization of full automation of the substation, and has certain risks. The inspection personnel need to observe at close range, and when the closing is not in place, the high-voltage disconnector generates heat, which may produce electric sparks or even explode.

[0005] It is extremely necessary to develop a double-checking method for automatically identifying the closing failure, and the State Grid and the Southern Power Grid have recently issued relevant notices. The current on-off state identification method is only a simple binary logic judgment, and the characteristics of the closing failure are not obvious. The closing position identification is more difficult than the on-off identification, and requires accurate identification devices and methods. The existing image acquisition or video monitoring method can only determine whether the switch is on or off, but cannot accurately determine whether the switch is closed. Other methods such as infrared method and attitude sensor are being studied. The infrared method needs to be powered for a period of time after closing, and the abnormal heating of the switch can be determined at this time. If the power is cut off at this time, it will cause certain economic losses and affect the normal operation of the power grid, which is a kind of after-the-fact remedy method. The attitude sensor needs to install equipment on the switch, and the power grid needs to be cut off for installation and maintenance, which affects the operation of the power grid. These methods have not achieved the expected results, and better double-checking methods need to be developed.

[0006] At present, there are various types and models of switch disconnectors in substations in China, and their working principles and appearances are quite different. The double-column horizontal rotary center break switch disconnector, hereinafter referred to as horizontal rotary switch disconnector, is one of the most widely used switch disconnectors in substations at present. The closing position state identification of this type of switch disconnector is studied in the present invention. SUMMARY

[0007] In order to overcome the defects in the prior art, the present invention provides a horizontal rotary switch disconnector state accurate identification device and a closing state judgment method to solve the above technical problems.

[0008] The present invention provides the following technical solutions:

[0009] A horizontal rotary switch disconnector state accurate identification device comprises:

[0010] A camera, a switch disconnector identification front-end device, a switch disconnector identification back-end learning and training device, and a data and control bus.

[0011] The camera is used to take images of the switch disconnector. The images are output to the switch disconnector identification front-end device through the data and control bus. The shooting action of the camera is controlled by the control signal sent by the switch disconnector identification front-end device.

[0012] The switch disconnector identification front-end device identifies the switch disconnector images and analyzes the included angle of the two rotary arms to determine whether the closing is in place. The signal of whether the closing is in place is sent to the switch disconnector identification back-end learning and training device. At the same time, the images of the switch disconnector are also forwarded to the switch disconnector identification back-end learning and training device. The identification parameters of the switch disconnector identification front-end device are provided and refreshed by the switch disconnector identification back-end learning and training device.

[0013] The knife switch recognition background learning and training device performs deep learning training according to the received image of the knife switch, and sends the learned and trained parameters to the knife switch recognition front-end device, and simultaneously forwards the signal of whether the closing is in place to the dispatching center, and multiple knife switch recognition front-end devices can share one knife switch recognition background learning and training device.

[0014] The data and control bus is used for transmitting data and control signals between the camera, the knife switch recognition front-end device and the knife switch recognition background learning and training device.

[0015] Preferably, the camera is arranged as follows: the bases of the three cameras are arranged in parallel and are installed in a straight line, and each camera is used to capture each phase of a three-phase knife switch, and the center of each camera is aligned with the contact joint of the two rotating arms of a knife switch.

[0016] A closing state judgment method using the horizontal rotating type knife switch state accurate recognition device, characterized in that, comprising the following steps:

[0017] Knife switch image acquisition and preprocessing: acquiring the image containing the knife switch from the camera, and performing image preprocessing;

[0018] Knife switch target recognition: finding the required knife switch in an image containing background information, filtering out the background, extracting the knife switch from the captured picture, and selecting the required knife switch to be detected with a smallest rectangular frame;

[0019] Edge extraction: extracting the contour lines of the two horizontal rotating arms of the knife switch within the range selected by the rectangular frame, so as to accurately detect the closing state by using a geometric method;

[0020] Linear detection of closing: calculating the error between the contour lines of the two horizontal rotating arms of the knife switch and linear 180 degrees, if the error exceeds a given threshold, a closing misplacement signal is sent, otherwise a closing success signal is sent, and if any one of the three-phase knife switches is not closed in place, the entire three-phase knife switch is regarded as not being closed in place.

[0021] Preferably, the knife switch image acquisition and preprocessing step comprises:

[0022] Data acquisition: issuing a shooting instruction to read the knife switch image captured by the camera;

[0023] Data cropping: different cameras have different resolutions, in order to adapt to different cameras and reduce unnecessary calculation, after obtaining the picture from the camera, the picture is uniformly cropped to 512*684 pixels;

[0024] Image noise removal: the filtered picture after cutting is filtered, and the value of each pixel point is obtained by weighted average of itself and other pixel values in the neighborhood. Each pixel in the image is scanned by a Gaussian kernel, and the weighted average gray value of the pixels in the field determined by the Gaussian kernel is calculated to replace the value of the convolution center pixel. The Gaussian kernel is an odd-sized Gaussian template.

[0025] Preferably, the target recognition step of the knife switch includes:

[0026] S1, the pre-processed picture is input into the pre-trained neural network to obtain the corresponding feature map;

[0027] S2, set ROI through each point in the feature map to obtain a plurality of ROI candidate boxes;

[0028] S3, the plurality of ROI candidate boxes are sent to RPN for binary classification and BB regression, and a part of the candidate ROIs are filtered out.

[0029] S4, the remaining ROIs are subjected to ROI Align operation, that is, the pixels of the original image and the feature map are first matched, and then the feature map and the fixed feature are matched.

[0030] S5, the ROIs are subjected to N-class classification, BB regression and mask generation.

[0031] Preferably, the edge extraction includes the following steps:

[0032] Constructing a HED network model;

[0033] Using the algorithm structure under multi-scale deep learning on the structure of the HED network model.

[0034] Preferably, the closing linear detection step includes:

[0035] Edge cutting: the image after edge extraction is further cut into a 256*256 size image, the number of pixel points is controlled within an acceptable range, the calculation time is saved, and the system reaction is accelerated.

[0036] Inclination direction determination: the inclination direction of the knife switch will affect the subsequent key pixel extraction. The knife switch image with left high and right low inclination angle is called left type, and vice versa. The inclination angle is negative, which is called right type. The judgment method is to traverse the pixel points in the upper half of each image, and respectively accumulate and sum the pixel point weights of the left and right parts. The left weight sum is higher, and the right weight sum is higher.

[0037] Key pixel extraction: the pixels of the upper half profile line of the left and right halves of the knife gate are called key pixels, and according to the positive and negative of the inclination angle, the irrelevant regions are further filtered out using the corresponding positive and negative diagonal lines, and finally the key pixels are obtained;

[0038] Pixel normalization: traverse the remaining regions in the picture after key pixel extraction, and assign 1 to the pixel points with a weight greater than the threshold value, and assign 0 to the points less than the threshold value;

[0039] Linear fitting and judgment: the least square method is used to linearly fit the key pixels to obtain a linear function, and then the linear function is used to check each key pixel to see whether it substantially conforms to the linear function, and finally the proportion of points conforming to the linear function to the key pixels is counted, which is called the correct rate. 70% correct rate is used as the qualified line, and greater than 70% is determined as the knife gate being completely closed, otherwise the gate is not closed in place.

[0040] The horizontal rotating knife gate state accurate identification device and the closing state judgment method provided by the application have the following advantages and beneficial effects compared with the prior art:

[0041] 1. It can accurately determine whether the horizontal rotating knife gate is closed in place;

[0042] 2. Good safety, non-contact optical video detection;

[0043] 3. No need to stop power installation, no need to stop power maintenance, no influence on the normal operation of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 It is a module composition structure and working principle schematic diagram of the horizontal rotating knife gate state accurate identification device of the embodiment of the application;

[0045] Figure 2 It is a flowchart of the closing state judgment method of the embodiment of the application;

[0046] Figure 3 It is an image after target recognition processing of the knife gate of the embodiment of the application;

[0047] Figure 4 It is an image after edge extraction processing of the knife gate of the embodiment of the application;

[0048] Figure 5 It is an image after linear detection of the closing of the embodiment of the application. DETAILED DESCRIPTION

[0049] The application will be further described in detail below in combination with the embodiments and the drawings, but the embodiments of the application are not limited thereto.

[0050] EMBODIMENT

[0051] Referring to Figures 1-5 The horizontal rotating knife switch state accurate identification device provided by the embodiment comprises a camera, a knife switch identification front-end device, a knife switch identification background learning and training device, a data and control bus, and a router.

[0052] Suppose that a certain transformer substation has n horizontal rotating knife switches, Figure 1 The left, middle and right three cameras for shooting the knife switch 1 and the left, middle and right three cameras for shooting the knife switch n are drawn, for the sake of simplicity, Figure 1 A dotted line is used to represent other knife switches. The bases of the left, middle and right three cameras are installed in a line, and each camera is used to shoot each phase of a three-phase knife switch. The center of each camera is aligned with the contact joint of the two rotating arms of a knife switch of each phase. The image of a knife switch of a certain phase of a three-phase horizontal rotating knife switch is as shown in Figure 3 The green box has two rotating arms, and the center is the closed contact. A three-phase knife switch has three such knife switches.

[0053] In Figure 1 A three-phase knife switch is equipped with a knife switch identification front-end device. When the knife switch identification front-end device sends a shooting signal, the knife switch images shot by the left, middle and right three cameras are output to the knife switch identification front-end device through the data and control bus.

[0054] The knife switch identification front-end device uses a deep learning algorithm to identify the knife switch image, analyzes the linear deviation value of the profile line of the two rotating arms, judges whether the closing is in place, and sends the signal of whether the closing is in place to the knife switch identification background learning and training device. At the same time, the image of the knife switch is also forwarded to the knife switch identification background learning and training device. The deep learning network parameters of the knife switch identification front-end device are provided and refreshed by the knife switch identification background learning and training device. Figure 1 In Figure 1 There are n three-phase knife switches, but only one knife switch identification background learning and training device is needed, because the deep learning network parameters of these knife switch identification front-end devices are the same. Further, multiple transformer substations can share one knife switch identification background learning and training device, so

[0055] The knife switch identification background learning and training device performs deep learning training according to the received knife switch image, sends the learned and trained parameters to the knife switch identification front-end device, and forwards the signal of whether the closing is in place to the dispatching center.

[0056] Figure 1The data and control bus in the horizontal rotating type switch state accurate identification device can adopt a general wired Ethernet or wireless wifi, all the cameras are network cameras, the switch identification front-end device and the switch identification background learning and training device have network interfaces, and data and control signals can be transmitted between any devices through the network. The dispatching center can also directly obtain the switch image to monitor the switch without the switch identification front-end device.

[0057] Figure 2 The closing state judgment method of the horizontal rotating type switch state accurate identification device.

[0058] The closing state judgment method of the horizontal rotating type switch state accurate identification device provided by the embodiment of the application specifically includes the following steps:

[0059] S1, the switch identification front-end device sends a shooting instruction to the left, middle and right three cameras, and the three cameras shoot a three-phase switch that has been closed together, and the cameras send the images of the three switches of the three-phase switch to the switch identification front-end device through the data and control bus.

[0060] S2, the switch identification front-end device calls an image preprocessing subprogram module to perform data cropping and noise removal on the image.

[0061] S3, the switch identification front-end device calls a switch target identification subprogram module to find the required switch in an image containing background information, filter out the background, extract the switch from the photographed picture, and select the required switch to be detected with a smallest rectangular frame, as shown in Figure 3 .

[0062] S4, the switch identification front-end device calls an edge extraction subprogram module to extract the contour lines of the two horizontal rotating arms of the switch within the range selected by the rectangular frame. The application adopts a HED network model, and then adopts an algorithm structure under multi-scale deep learning on the HED network model structure.

[0063] S5, the switch identification front-end device calls a closing linear detection subprogram module to calculate the error of the contour lines of the two horizontal rotating arms of the switch and the linearity of 180 degrees. If the error exceeds a given threshold value, a closing misplacement signal is sent, otherwise a closing success signal is sent. If any one of the three-phase switches is not closed in place, the entire three-phase switch is classified as not closed in place.

[0064] S6, the switch identification background learning and training device collects all the closing judgment signals of the switch identification front-end devices, presents through an interface, and forwards to the dispatching center.

[0065] In the Figure 2 flow, the specific steps of the image preprocessing subprogram module are as follows:

[0066] Data clipping: different camera resolutions are different, in order to adapt to different cameras while reducing unnecessary calculation, after obtaining the photos from the camera, the pictures are uniformly clipped to 512*684 pixels;

[0067] Image noise removal: the filtered picture is filtered by using a known filtering algorithm, the weighted average of the whole image is calculated, the value of each pixel point is obtained by weighted average of itself and other pixel values in the neighborhood, each pixel in the image is scanned by using a Gaussian kernel, and the weighted average gray value of the pixels in the field determined by the Gaussian kernel is calculated, and the value of the convolution center pixel is replaced by the weighted average gray value, the Gaussian kernel is an odd size Gaussian template, and a 5*5 Gaussian template is used here.

[0068] The above Figure 2 In the process, the specific steps of the knife switch target identification subprogram module are as follows:

[0069] S1, the preprocessed picture is transmitted into the pre-trained neural network to obtain the corresponding feature map;

[0070] S2, set ROI through each point in the feature map to obtain a plurality of ROI candidate boxes;

[0071] S3, the plurality of ROI candidate boxes are sent to RPN for binary classification and BB regression, and a part of the candidate ROIs are filtered out;

[0072] S4, the remaining ROIs are subjected to ROI Align operation, that is, the pixels of the original picture and the feature map are first corresponded, and then the feature map and the fixed feature are corresponded;

[0073] S5, the ROIs are subjected to N-class classification, BB regression and mask generation.

[0074] Figure 3 The green box is obtained by the knife switch target identification subprogram, and more details can be found in the Mask R-CNN algorithm.

[0075] Figure 4 Obtained by the edge extraction subprogram module, see the HED algorithm.

[0076] The above Figure 2 In the process, the specific steps of the closing linear detection subprogram module are as follows:

[0077] Edge clipping: the edge extracted image is further clipped to 256*256 size image, the number of pixel points is controlled within an acceptable range, the calculation time is saved, and the system reaction is accelerated;

[0078] Inclination direction determination: the direction of the inclination of the knife switch will affect the subsequent key pixel extraction. The knife switch image with the left high and right low inclination angle is called left type, and the knife switch image with the negative inclination angle is called right type. The determination method is to traverse the pixel points in the upper half of each image, and to accumulate and sum the pixel point weights of the left and right parts, respectively. The left weight sum is higher, and the right weight sum is higher.

[0079] Key pixel extraction: the pixels of the upper half of the profile line of the left and right halves of the knife switch are called key pixels. According to the positive and negative of the inclination angle, the corresponding positive and negative diagonal lines are used to further filter out irrelevant areas, and finally the key pixels are obtained.

[0080] Pixel point normalization: the remaining area in the picture after key pixel extraction is traversed, and the pixel point weight greater than the threshold is assigned 1, and the point less than the threshold is set to 0.

[0081] Linear fitting and determination: the least square method is used to linearly fit the key pixels, a linear function is obtained, and the linear function is used to check each key pixel to see whether it substantially conforms to the linear function. Finally, the proportion of points conforming to the linear function to the key pixels is counted, and is recorded as the correct rate. A qualified line of 70% correct rate is used as a criterion. If the correct rate is greater than 70%, the knife switch is determined to be completely closed, otherwise the knife switch is not closed in place. The final result is shown in Figure 5 .

[0082] Figure 1 The knife switch recognition background learning and training device includes a deep learning server as hardware, and installs the deep learning software related to the knife switch in the application, including knife switch target recognition and edge extraction. The knife switch recognition background learning and training device performs deep learning training according to the received knife switch image, and sends the learned and trained parameters to the knife switch recognition front-end device. Without these parameters, the corresponding program in the knife switch recognition front-end device cannot run correctly.

[0083] Another function of the knife switch recognition background learning and training device is to collect the signals of whether the knife switch is in place from multiple knife switch recognition front-end devices, present the signals through an interface, and forward the signals to a dispatching center. Multiple knife switch recognition front-end devices in a substation can share one knife switch recognition background learning and training device, and even multiple substations can share one knife switch recognition background learning and training device.

[0084] The method and device provided by the above embodiment of the application adopt a cloud-edge fusion architecture, target recognition is performed by using a Mask R-CNN network, edge extraction is performed by using an HED method, and a method for detecting that a closing switch is not in place is proposed on this basis, so that accurate recognition of the state of the knife switch is realized, manual inspection can be replaced, manpower is saved, and safety is improved.

[0085] The above merely illustrates the embodiments of the present application, and is not used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for judging the closing state of a horizontal rotary knife switch state accurate identification device, characterized in that, It comprises the following steps: Knife switch image acquisition and preprocessing: obtain the image containing the knife switch from the camera, and perform image preprocessing, the camera is arranged as follows: The bases of the three cameras are installed in a line side by side, and each camera respectively captures each phase knife switch of a three-phase knife switch, the center of each camera is respectively aligned with the contact joint of the two rotating arms of a knife switch, the knife switch is a single-phase knife switch, and a three-phase knife switch comprises three single-phase knife switches connected in series; Knife switch target recognition: find the required knife switch in an image containing background information, filter out the background, and extract the knife switch from the captured picture, and select the required knife switch to be detected with a smallest rectangular frame; Edge extraction: the contour lines of the two horizontal rotating arms of the knife switch are extracted within the range selected by the rectangular frame, so that the closing state is accurately detected by using a geometric method; Linear detection of closing: calculate the error between the contour lines of the two horizontal rotating arms of the knife switch and linear 180 degrees, if the error exceeds a given threshold, a closing misalignment signal is sent, otherwise a closing success signal is sent, if any one of the three-phase knife switches is not closed, the whole three-phase knife switch is classified as closing misalignment, and the steps specifically include: Edge clipping: the image after edge extraction is further clipped to an image with a size of 256*256, the number of pixel points is controlled within an acceptable range, the calculation time is saved, and the system response is accelerated; Inclination direction judgment: the inclination direction of the knife switch will affect the subsequent key pixel extraction, the knife switch image with the inclination angle being high on the left and low on the right is called left type, and vice versa, the inclination angle is negative, which is called right type, the judgment method is to traverse the pixel points in the upper half of each image, and the pixel point weights on the left and right are respectively accumulated and summed, the left weight sum is higher, and the right weight sum is higher; Key pixel extraction: the pixels of the upper half contour line of the left and right halves of the knife switch are called key pixels, according to the positive and negative of the inclination angle, the corresponding positive and negative diagonal lines are used to further filter out irrelevant areas, and finally the key pixels are obtained; Pixel point normalization: traverse the remaining area in the picture after key pixel extraction, and assign 1 to the pixel point weight greater than the threshold value, and assign 0 to the pixel point weight less than the threshold value; Linear fitting and judgment: the least square method is used for linear fitting of the key pixels, a linear function is obtained, and the linear function is used to test each key pixel to see whether it substantially conforms to the linear function, and finally the proportion of the points conforming to the linear function to the key pixels is counted, which is called the correct rate; The qualified line is 70% correct rate, greater than 70% is determined as completely closed, otherwise it is closing misalignment.

2. The closing state determination method according to claim 1, characterized by The knife switch image acquisition and preprocessing step comprises: Data acquisition: send a shooting instruction and read the knife switch image captured by the camera; Data clipping: different cameras have different resolutions, in order to adapt to different cameras and reduce unnecessary calculation, after obtaining the picture from the camera, the picture is uniformly clipped to a 512*684 pixel picture; Image noise removal: a filtering algorithm is used to filter the cropped picture, and the value of each pixel point is obtained by weighted average of itself and other pixel values in the neighborhood. A Gaussian kernel is used to scan each pixel in the image, and the weighted average gray value of the pixels in the domain determined by the Gaussian kernel is calculated and used to replace the value of the convolution center pixel. The Gaussian kernel is an odd-sized Gaussian template.

3. The closing state determination method according to claim 1, characterized by, The target recognition step of the knife switch includes: S1, input the preprocessed picture into the pre-trained neural network to obtain the corresponding feature map; S2, set ROI through each point in the feature map to obtain a plurality of ROI candidate boxes; S3, send the plurality of ROI candidate boxes to RPN for binary classification and BB regression, and filter out a part of the candidate ROIs; S4, perform ROI Align operation on the remaining ROIs, that is, first correspond the pixels of the original image and the feature map, and then correspond the feature map and the fixed feature; S5, perform N-class classification, BB regression and mask generation on the ROIs.

4. The closing state determination method according to claim 1, characterized by, The edge extraction step includes: Constructing a HED network model; Using the algorithm structure under multi-scale deep learning on the HED network model structure.

5. A horizontal rotary knife gate state accurate recognition device, characterized by, A closing state judgment method according to any one of claims 1 to 4, comprising: a camera, a knife switch recognition front-end device, a knife switch recognition back-end learning and training device, and a data and control bus; The camera is used to shoot the image of the knife switch, and the image is output to the knife switch recognition front-end device through the data and control bus. The shooting action of the camera is controlled by the control signal sent by the knife switch recognition front-end device; The knife switch recognition front-end device identifies the knife switch image and analyzes the included angle of the two rotating arms to determine whether the closing is in place, and sends the signal of whether the closing is in place to the knife switch recognition back-end learning and training device. The image of the knife switch is also forwarded to the knife switch recognition back-end learning and training device. The identification parameters of the knife switch recognition front-end device are provided and refreshed by the knife switch recognition back-end learning and training device; The knife switch recognition back-end learning and training device performs deep learning training according to the received knife switch image, and sends the learned and trained parameters to the knife switch recognition front-end device. The signal of whether the closing is in place is also forwarded to the dispatching center. Multiple knife switch recognition front-end devices can share one knife switch recognition back-end learning and training device; The data and control bus is used to transmit data and control signals between the camera, the knife switch recognition front-end device and the knife switch recognition back-end learning and training device.

Images