A photovoltaic fault detection method based on unmanned aerial vehicle cruising

Abstract

The application discloses a photovoltaic fault detection method based on unmanned aerial vehicle (UAV) cruising, and belongs to the field of pattern recognition, and comprises the following steps: acquiring high-definition true color images by UAV adaptive cruising; synthesizing images to acquire a high-definition general image of a whole station; splitting each acquired image; establishing a detection model; identifying fault causes by using the detection model, and outputting fault frame coordinates containing fault images; pixel positioning of the fault frame; and marking the identified position on the high-definition general image. According to the technical scheme, the UAV can automatically change the shooting height and angle to ensure the consistency of the acquired images; the high-definition general image of the whole station can be acquired through image synthesis; the algorithm efficiency can be improved under the premise of ensuring the accuracy through image splitting; various visible fault types can be found, the hot spots of the photovoltaic panel can be cleaned in time, the loss of the photovoltaic module can be reduced, and the fault position can be accurately determined through pixel-level GPS positioning calculation.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition, specifically relating to a method for identifying and locating faults in distributed photovoltaic modules by acquiring high-definition true-color images using drones. Background Technology

[0002] With the increasingly widespread application of photovoltaic power generation technology, a large number of photovoltaic systems have been put into operation. The actual operating environment of photovoltaic power plants is harsh, and the photovoltaic array faults are complex and diverse, which seriously affects the safe and stable operation and economic benefits of the power plants.

[0003] Major failures in photovoltaic (PV) systems include microcracks in PV modules before and during operation; hot spots caused by localized shading or uneven sunlight reception; performance degradation of PV modules due to potential-induced degradation and corrosion; array mismatch caused by performance degradation of some modules or improper connection and arrangement; and open circuits or short circuits in PV modules caused by incorrect junction boxes. These failures severely reduce the lifespan of PV modules and various electrical equipment, reduce the power generation of PV power plants, and even seriously affect the safe operation of the entire PV system.

[0004] Currently, the main methods for fault diagnosis in photovoltaic power generation include methods based on circuit structure, methods based on infrared images, methods based on electrical measurements, methods based on mathematical reference models, and fault diagnosis methods based on intelligent detection.

[0005] Photovoltaic (PV) faults can be broadly categorized into two types: 1. Visible faults, such as bird droppings, leaves, dust, etc., which are visible obstructions; 2. Hot spot issues. The first type of fault can be resolved by cleaning the surface of the PV panel, while the second type requires replacing the PV module.

[0006] Existing technologies mainly focus on the identification of hot spot faults, which are difficult to detect with the naked eye. They generally use infrared imaging, such as Chinese patent application CN115100296A.

[0007] Faults such as prolonged partial shading of photovoltaic panels or the surface being covered by foreign objects can cause hot spots if not detected and cleaned in time, resulting in more serious damage to the photovoltaic modules. Summary of the Invention

[0008] To address the aforementioned problems, this invention is proposed. This invention employs the following technical means: a photovoltaic fault detection method based on unmanned aerial vehicle (UAV) patrol, comprising the following steps: 1.1. UAV adaptive cruise acquires high-definition true-color images; 1.2. Based on the GPS information obtained from the acquired images, image synthesis is performed to obtain a high-definition overall map of the entire station; 1.3. Segment each acquired image; 1.4. Train, validate, and test the detection model for different causes of failure; 1.5. Use the detection model to identify the cause of the fault and output the coordinates of the fault box containing the fault image; 1.6 Fault box pixel location; 1.7. Based on image GPS information and pixel positioning, mark and identify locations on the high-definition overall map.

[0009] This application uses drones to capture high-definition true-color images, synthesizes a high-definition overall map of the entire site, which is convenient for display and marking, and uses high-definition true-color images to build and identify models of faults that can be identified by the naked eye, thus achieving the purpose of the invention.

[0010] Beneficial effects: The technical solution proposed in this invention enables UAV adaptive cruise, automatically changing the shooting altitude and angle to ensure the consistency of the acquired images; high-definition overall map of the entire site is obtained through image synthesis; the algorithm efficiency is improved while ensuring accuracy through image segmentation; various types of faults visible to the naked eye can be detected, and timely cleaning can prevent hot spots from forming on photovoltaic panels and reduce the loss of photovoltaic modules; pixel-level GPS positioning calculation accurately determines the fault location. Attached Figure Description

[0011] Figure 1 The coverage area of ​​a single image. Figure 2 This is a partial site plan. Figure 3 The image is faulty. Figure 4 A schematic diagram illustrating the relationship between the fault box, the split diagram, and the original diagram. Detailed Implementation

[0012] A photovoltaic fault detection method based on UAV patrol includes the following steps: 1.1. UAV adaptive cruise acquires high-definition true-color images; 1.2. Based on the GPS information obtained from the acquired images, image synthesis is performed to obtain a high-definition overall map of the entire station; 1.3. Segment each acquired image; 1.4. Train, validate, and test the detection model for different causes of failure; 1.5. Use the detection model to identify the cause of the fault and locate the pixel; 1.6. Based on image GPS information and pixel positioning, mark and identify locations on the high-definition master map.

[0013] First, use drone adaptive cruise to acquire high-definition true-color images.

[0014] Before shooting, set the drone inspection parameters: Based on the position and tilt angle of the photovoltaic panels, set the drone's flight parameters to ensure that the shooting height and angle for all photovoltaic panels are roughly consistent. The drone inspection parameters include the drone flight path, the drone camera gimbal angle, the shooting mode, and the angle and direction. In this embodiment, the camera is positioned at a height of 15-20 meters. Photovoltaic panels are generally laid at a certain angle to the ground, and the angle varies depending on the region's lighting conditions; therefore, image acquisition is generally done by flying perpendicular to the ground. Save the flight parameters so that the drone can fly according to the predetermined route and parameters during project use, eliminating the need to reset them each time.

[0015] Images captured by drones may have deviations when directly mapped to the ground due to factors such as shooting angle. In order to accurately map the images to the actual physical location coordinate system (latitude and longitude coordinate system), in this embodiment, before synthesizing the images, the high-definition true-color images acquired by the drone's adaptive cruise are transformed by perspective, and the projection position of the images on the plane is calculated.

[0016] Viewpoint transformation includes: Camera transformation: Constructing a right-handed coordinate system with the drone as the origin and the camera's orientation as the x-axis, and using matrix transformation to convert the world coordinate system into the camera coordinate system; Projection transformation: Projecting the 3D images captured by the camera into a 2D space using a projection transformation matrix; Viewport transformation: Transforming the images in the 2D space of the drone into digital images in the actual physical world using a back projection matrix.

[0017] The above processing is existing technology and will not be elaborated on here.

[0018] In this embodiment, the above transformation processing is only applied to the images used for image synthesis to obtain a high-definition overall map of the entire station. The original images taken by the UAV are used for model training and fault identification.

[0019] Image synthesis is performed based on GPS information from the acquired images to obtain a high-resolution overall map of the entire site.

[0020] The synthesis of a high-resolution overall site map involves combining digital images into a standard OSM (OpenStreetMap) map. In this embodiment, the geographical coordinates of the site, i.e., the latitude and longitude range, are first determined to obtain the size of the coverage area. Then, based on the coverage area of ​​a single photo acquired by the UAV, the entire site is filled with photos. The pixel size of the synthesized image is calculated according to the highest resolution of the photos. On this basis, the photos are synthesized according to latitude and longitude to obtain the correspondence between any point in the geographical coordinate area and the pixel coordinates.

[0021] Specifically, the following steps are included: First, determine the height h and width w of the composite image, which is achieved through the following steps.

[0022] 4.1.1 Determine the latitude and longitude range (maxLon, minLon, maxLat, minLat) of the entire station's geographical coordinate area, where maxLon is the maximum longitude value, minLon is the minimum longitude value, maxLat is the maximum latitude value, and minLat is the minimum latitude value.

[0023] 4.1.2 Obtain the pixel values ​​Sh*Sw of a single image, where Sh is the pixel value of the height of the single image and Sw is the pixel value of the width of the single image; obtain the coverage area Ch*Cw of a single image, where Ch is the height of the coverage area of ​​the single image in meters and Cw is the width of the coverage area of ​​the single image in meters.

[0024] If a single image has a pixel value (i.e., resolution) of 5472*3648, at a cruising altitude of 15 meters, each image will show 5-6 rows of 18 photovoltaic modules per row, covering an area of ​​approximately 18*12m. 2 ,like Figure 1 As shown.

[0025] 4.1.3 Calculate the number of images required to cover the entire station with a single image: Rw=(maxLon-minLon) / Cw, Rh=(maxLat-minLat) / Ch, where maxLon-minLon and maxLat-minLat are converted to meters.

[0026] Rw represents the number of images required for coverage in the east-west direction, and Rh represents the number of images required for coverage in the north-south direction.

[0027] The above calculation formula is just an illustration.

[0028] At the equator, a 1° difference in longitude corresponds to a distance difference of approximately 111 kilometers; at any latitude, a 1° difference in longitude corresponds to a distance difference of approximately 111*cosα kilometers, where α represents the latitude.

[0029] A 1° difference in latitude along the same meridian is approximately 111 kilometers. For every 0.00001 degrees of latitude, the distance differs by approximately 1.1 meters; for every 0.0001 degrees, the distance differs by approximately 11 meters; for every 0.001 degrees, the distance differs by approximately 111 meters; for every 0.01 degrees, the distance differs by approximately 1113 meters; and for every 0.1 degrees, the distance differs by approximately 11132 meters.

[0030] Based on the above, precise calculations can be performed.

[0031] 4.1.4 Calculate the height h and width w of the synthesized image, h = Sh * Rh, w = Sw * Rw, where h and w are pixel values.

[0032] The height h corresponds to latitude, i.e., the north-south direction, and the width w corresponds to longitude, i.e., the east-west direction.

[0033] The latitude and longitude of each image are obtained from the GPS information. The first step of image synthesis is performed in the latitude and longitude space, and the duplicates are removed from the overlapping areas.

[0034] Given the GPS center point of the image, the flight attitude of the UAV, and the camera parameters, a digital image is obtained through transformation. The features and positions of similar objects in adjacent images are identified, and the images are transformed and stitched together. Two images with partially overlapping areas are stitched together into one image, and so on, all images are combined to form a high-definition large image.

[0035] The correspondence between any point in a geographic coordinate region and its pixel coordinates: X = (lon - minLon) * w / (maxLon-minLon), Y = (maxLat - lat) * h / (maxLat-minLat), Where (lon, lat) are the longitude and latitude of a point in the geographic coordinate system, and (X, Y) are the corresponding pixel coordinates.

[0036] The relative and absolute positions of the images are calculated based on the screen coordinate information. Each image is then placed in its corresponding grid using a raster method to complete the image composition.

[0037] Within the site, there may only be a few areas with photovoltaic modules. In order to reduce the workload, in this embodiment, the drone only takes pictures of the areas where photovoltaic modules are installed. After the high-definition image is synthesized, it is combined with the existing electronic map to generate the final site map.

[0038] Existing electronic maps include buildings, roads, and terrain features. Combining these two types of maps creates a high-resolution image that overlays the corresponding latitude and longitude range of the electronic map, presenting more information in a more intuitive way.

[0039] The site master plan was obtained through the above steps. Figure 2 It is part of the site master plan.

[0040] Each acquired image is split into its constituent parts.

[0041] In daily use, drones are used to periodically patrol and acquire high-definition true-color images for fault diagnosis.

[0042] The images acquired by the drone have a relatively high resolution, and using the entire image directly for algorithm recognition would affect efficiency. In this embodiment, the image is split into multiple images of the optimal resolution accepted by the algorithm, achieving faster computation speed.

[0043] Define the size of the split image as Dh*Dw, where Dh and Dw are the pixel values ​​of the height and width of the split image, respectively; define the size of the overlapping region as m%; based on the above parameters, use the sliding window technique to split the image to obtain multiple split images.

[0044] Since the number of pixels in the original image is not necessarily an integer multiple of Dh and Dw, it is necessary to set the overlap area m%, that is, the horizontal and / or vertical overlap of adjacent split images is m%, where m is between 10 and 20.

[0045] In this embodiment, Dh and Dw have the same value, preferably 1280 or 960.

[0046] The sliding window technique described above uses a Dh*Dw window to slide and position itself on the original image according to the repeating regions, taking one image at each position.

[0047] A specific implementation example is as follows: (1) Define the size of the split image. Here, the width and height of the split image are set to the same value, which is Wp.

[0048] (2) Define the size of the overlapping area. In order to ensure that the target to be detected is not split, the width or height of the split image needs to be set to the overlapping area. The coverage area is overlay=Wp*m.

[0049] (3) The synthesized image is split into multiple split images based on the sliding window technique. The sliding step size is step. Then the relationship between W, step, and overlay is: Wp ≤ step + overlay .

[0050] Taking width as an example, if the width of the original image is greater than the W value, it needs to be split cyclically based on the width: When the original image width W origin ≥ step * index + overlay At that time, according to the x-coordinate ( step * (index - 1), step * index + overlay ( ) to be cut and split. W origin The original image width, index This represents the number of images currently split.

[0051] The height splitting method is the same as the width splitting method. Simultaneously with obtaining the split image, the position information of the split image is also obtained, namely the pixel coordinates (x_min, y_max) and (x_max, y_min) of the top-left and bottom-right corners of the split image in the original image.

[0052] The positional relationship information between the split image and the original image will be cached in file format for subsequent calculations.

[0053] Train, validate, and test the detection model for different failure reasons.

[0054] 1. Dataset preparation.

[0055] Obtain the photovoltaic fault dataset FP, and divide the photovoltaic fault dataset into a training dataset TRFP and a test dataset TEFP.

[0056] Randomly select 80% of the TRFP and TEFP datasets for training and 20% for testing.

[0057] In this invention, different fault types are set as bird droppings, leaves, and other foreign objects.

[0058] 2. Organize data training.

[0059] Training data generation method: Use the labelImg tool for image annotation to generate txt format labels, and each line in the file represents the fault box information.

[0060] Network structure: yolov5 model, and the calculation methods of Conv, activation function, and gradient update are as follows: The calculation method of Conv is as follows: , where the input feature map I(n,c) R H×W , the filter kernel W(k,c) R R×S , the output feature map O(n,k) R P×Q , n < N, K < k, N, C, and K respectively represent the batch size, input channels, and output channels.

[0061] The calculation method of the activation function Sigmoid is as follows: , where x represents the input, S ( x ) represents the output.

[0062] The parameter update formula of SGD is: θ t+1 = θ t –α t f it (θt )， in it {1,2,…,n} Let represent the sample index randomly drawn according to a uniform distribution in the t-th iteration, where the stochastic gradient of one sample is . f it (θ t ) , α t Let be the learning rate for the t-th iteration, used to adjust the magnitude of parameter updates. θ t Let be the target parameters for round t.

[0063] 3. Organize data testing.

[0064] The test data is generated in the same way as the training data, and the model testing method is based on existing technology, which will not be described in detail here.

[0065] The detection model is used to identify the cause of the fault and locate the pixel.

[0066] After the images acquired by the drone are segmented, a detection model is used to identify the cause of the fault, and then the fault location is located pixel by pixel.

[0067] After image segmentation, the fault was located in the segmented image.

[0068] After fault identification, the algorithm outputs fault box information, which contains the identified fault image and includes coordinates in YOLO format.

[0069] Fault box information includes: label_index, percent x , percent y , percent w , percent h ， confidence .

[0070] in label_index The type of object being identified is the fault type. In this embodiment, 1 represents bird droppings, 2 represents leaves, and 3 represents other foreign objects. [[ID=)26]]percent x The x-coordinate of the fault bounding box center point relative to the pixel width of the split image. percent y The ratio of the pixel coordinate y of the fault bounding box center point to the pixel height of the split image. percent wThis is the ratio of the pixel width of the fault bounding box to the pixel width of the split image. percent h This represents the ratio of the pixel height of the fault bounding box to the pixel height of the split image. confidence This represents the confidence level, with a maximum value of 100%, indicating the confidence level of detecting a fault. confidence > 92%, The fault was detected.

[0071] like Figure 4 As shown, the detection model outputs coordinates in YOLO format. Since fault identification is based on the split image, the coordinates need to be restored to their original coordinates in the image.

[0072] 1. Analyze the output file identified by the algorithm and calculate the location of the fault box in the original image.

[0073] 1-1. First, calculate the YOLO format coordinates of the identified split image in the original image.

[0074] , , , , in, W origin The original image pixel width, H origin In this embodiment, the original image pixel height is used. W origin = H origin , (x_min, y_max) are the pixel coordinates of the top left and bottom right corners of the split image in the original image.

[0075] Percent x_split This represents the ratio of the x-pixel coordinate of the split image's center point to the pixel width of the original image. Percent y_split To split the image, the ratio of the y-pixel coordinate of the center point to the pixel height of the original image. Percent w-split This represents the ratio of the pixel width of the split image to the pixel width of the original image. Percent h-split This represents the ratio of the pixel height of the split image to the pixel height of the original image.

[0076] 1-2. Calculate the YOLO format coordinates of the fault box in the original image.

[0077] , , , , 2. Calculate the top-left and bottom-right coordinates of the fault box containing the faulty object.

[0078] The coordinates of the top left corner of the fault box ( fault x-left , fault y-top )for: fault x-left =W origin *percent x_new –W origin *percent w_new *0.5， fault y-top =H origin y_new +H origin *percent h_new *0.5。

[0079] The coordinates of the lower right corner of the fault box ( *percent x-right fault y-bottom )for: , fault x-right =W origin fault x_new +W origin *percent w_new *0.5， *percent y-bottom =H origin fault y_new –H origin *percent h_new *0.5。

[0080] 3. Parse the Exif information of the original image and calculate the latitude and longitude of the fault box.

[0081] Read the image's Exif information and obtain the latitude and longitude coordinates of the image's center pixel.*percent longtitude and Center latitude The image was captured at a relative flight altitude of Center altitude The camera's calibrated focal length is Center focal length is CalibratedFocalLength (Unit: mm), the image calibration center point is located at ( FocalLength x CalibratedCenter y The yaw angle during camera shooting is , CalibratedCenter Calculate the pixel position of any point on the image ( yaw The relative latitude and longitude distance from the center point is: , , Using the hadersine package in Python, input the latitude and longitude coordinates of the center pixel of the image. Px, Py longtitude and Center latitude Yaw angle during camera shooting Center The relative distance calculated above yields the latitude and longitude of the target pixel location in the Gaussian projection. P longtitude ，P latitude ).

[0082] To further correct the deviation, this embodiment also uses the calculated ( P longtitude ，P latitude The geodetic GPS coordinates are calculated using the Gaussian projection inverse calculation formula.

[0083] First, calculate using the following formula. B f , B f This is the latitude of the base point calculated from the arc length of the meridian.

[0084] .

[0085] in, , X This is the arc length of the prime meridian. x for P latitude , y for P longtitude , L0 To designate the longitude of the central meridian, we set it at 120 degrees east longitude in the East Eighth Time Zone.a For the semi-major axis of the Earth's rotational ellipsoid, b For the short half-axis, e Here, it represents the eccentricity of the ellipse. .

[0086] Calculate geodetic GPS coordinates using the Gaussian projection inverse calculation formula: .

[0087] .

[0088] in, , , , .

[0089] The previously obtained ( yaw x-left fault y-top (), , fault x-right fault y-bottom Substitute them into the above calculation method respectively ( , fault ), thus obtaining the latitude and longitude region of the fault box on the Earth's surface ( Px, Py longtitude1 ， faul latitude1 (), fault longtitude2 faul latitude2 ).

[0090] All fault point latitude and longitude information, fault type, and fault confidence information are transformed into arrays, and then input into the NMS method of the PyTorch framework for deduplication to obtain fault detection data without duplicates.

[0091] Based on the above calculations, the latitude and longitude coordinates of the fault box are calculated according to the pixel coordinates of the image. Furthermore, the geodetic GPS coordinates are calculated using the latitude and longitude coordinates. Using the functions of map software, the area where the fault is located can be accurately marked on the high-definition master map of the entire site.

[0092] To visualize the specific fault type on the overall site map, after determining the fault's latitude and longitude information, fault photos are overlaid onto the high-resolution overall site map, such as... , fault Figure 3 As shown, the gray box indicates the fault location. After the fault is eliminated, manually remove the fault photos and restore the overall layout.

[0093] The photos taken daily are used for fault identification and location calculation, and do not need to be saved after completion.

Claims

1. A photovoltaic fault detection method based on unmanned aerial vehicle (UAV) patrol, characterized in that, Includes the following steps: 1.

1. UAV adaptive cruise acquires high-definition true-color images; 1.

2. Based on the GPS information obtained from the acquired images, image synthesis is performed to obtain a high-definition overall map of the entire station; 1.

3. Segment each acquired image; 1.

4. Train, validate, and test the detection model for different causes of failure; 1.

5. Use the detection model to identify the cause of the fault and output the coordinates of the fault box containing the fault image; 1.6 Fault box pixel location; 1.

7. Based on image GPS information and pixel positioning, mark and identify locations on the high-definition overall map; In step 1.2, after the UAV acquires high-definition true-color images through adaptive cruise, it first performs a viewpoint transformation on the images and calculates the projection position of the images on the plane. The viewpoint transformation includes: camera transformation: constructing a right-handed coordinate system with the UAV as the origin and the camera orientation as the x-axis, and using matrix transformation to convert the world coordinate system into the camera coordinate system; projection transformation: projecting the 3D image captured by the camera into a 2D space using a projection transformation matrix; viewport transformation: transforming the image in the 2D space into a digital image in the actual physical world using a back projection matrix. The image synthesis described in step 1.2 involves synthesizing digital images into a standard OSM map, including: 4.1 Determine the height h and width w of the composite image; 4.2 Obtain the latitude and longitude from the GPS information of each image; 4.

3. Perform image synthesis in latitude and longitude space, and remove duplicates from overlapping areas; 4.4 Obtain the correspondence between any point in the geographic coordinate region and pixel coordinates: X = (lon - minLon) * w / (maxLon-minLon), Y = (maxLat - lat) * h / (maxLat-minLat), Where (lon, lat) are the longitude and latitude in the geographic coordinate system, and (X, Y) are the corresponding pixel coordinates; 4.

5. Using the raster method, place each image into its corresponding grid to complete the image composite; Step 4.1 includes: 4.1.1 Determine the latitude and longitude range (maxLon, minLon, maxLat, minLat) of the entire station's geographical coordinate area, where maxLon is the maximum longitude value, minLon is the minimum longitude value, maxLat is the maximum latitude value, and minLat is the minimum latitude value. 4.1.

2. Obtain the pixel values ​​Sh*Sw of a single image, where Sh is the pixel value of the height of the single image and Sw is the pixel value of the width of the single image; obtain the coverage area Ch*Cw of a single image, where Ch is the height of the coverage area of ​​the single image in meters and Cw is the width of the coverage area of ​​the single image in meters. 4.1.3 Calculate Rw = (maxLon - minLon) / Cw, Rh = (maxLat - minLat) / Ch. Where maxLon-minLon and maxLat-minLat are converted to meters; 4.1.4 Calculate the height h and width w of the synthesized image, h = Sh * Rh, w = Sw * Rw, where h and w are pixel values.

2. The photovoltaic fault detection method according to claim 1, characterized in that, In step 1.1, the drone inspection parameters are set, including the drone flight path, drone camera gimbal angle, shooting mode, angle and direction.

3. The photovoltaic fault detection method according to claim 1, characterized in that, The high-resolution image of the entire site will be synthesized and then combined with the existing electronic map.

4. The photovoltaic fault detection method according to claim 1, characterized in that, 1.3 includes: Define the size of the split image as Dh*Dw, where Dh and Dw are the pixel values ​​of the height and width of the split image, respectively; Define the size of the overlapping region as m%; Based on the above parameters, the image is split using a sliding window technique to obtain multiple split images.

5. The photovoltaic fault detection method according to claim 4, characterized in that, 1.4 includes: Dataset preparation: Obtain the dataset FP and divide the dataset FP into the training dataset TRFP and the test dataset TEFP; Training data organization: Fault types are set as bird droppings, leaves and other foreign objects; Training data generation method: Image annotation is performed using the labelImg tool to generate txt format labels, with each line in the file representing fault box information; The YOLOv5 model is used for training and testing to generate a detection model.

6. The photovoltaic fault detection method according to claim 5, characterized in that, The fault box pixel localization described in step 1.6 includes: 1.6.

1. Parse the output file identified by the algorithm and calculate the location of the fault box in the original image; 1.6.2 Calculate the upper left and lower right coordinates of the fault box; 1.6.

3. Parse the Exif information of the original image and calculate the latitude and longitude position of the fault box.

7. The photovoltaic fault detection method according to claim 6, characterized in that, After obtaining information on all fault points, the latitude and longitude information, fault type, and fault confidence information of all fault points are transformed into arrays, which are then input into the NMS method of the PyTorch framework for deduplication to obtain fault detection data without duplicates.

8. The photovoltaic fault detection method according to claim 1, characterized in that, In step 1.7, photos containing faults will also be overlaid onto the overall high-definition map of the entire station.

Images