Photovoltaic panel defect detection system and method based on unmanned aerial vehicle and machine vision

Abstract

The invention discloses a photovoltaic panel defect detection system and method based on unmanned aerial vehicle and machine vision, the system comprises an unmanned aerial vehicle, a visible light lens and an EL lens, the EL lens is connected with the input end of an image collection module through a wireless network, and the output end of the image collection module is connected with the input end of an image processing module; the output end of the central processor is connected with the defect discrimination module; the unmanned aerial vehicle firstly carries a visible light lens and is used for automatically planning a route according to a given area range to carry out primary high-altitude aerial photography and simulated mesoscale aerial photography; secondly, carrying an EL lens, and carrying out multi-view adaptive directional high-precision shooting according to a route planning path; the image processing module is used for correcting a photovoltaic panel image collected by an EL lens, defect features are extracted through a SimAM-OfficientNetV2 network model, a sample set is classified into a defect-free sample and a defect sample, high-precision real-time detection is carried out on various defects existing in the image through a YOLOV9 network model, and rapid discrimination of the defects is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic panel defect detection, and specifically relates to a photovoltaic panel defect detection system and method based on unmanned aerial vehicle and machine vision. Background Art

[0002] Photovoltaic cells are one of the important products of the photovoltaic industry and have a broad prospect. However, when photovoltaic cells are produced and processed or used in photovoltaic stations, various defects such as missing corners, cracks, black spots and black cores will inevitably appear during the production process or after long-term use. These defects seriously affect the performance and stability of the cell. Therefore, it is very important to perform defect detection on photovoltaic cells during the production process or when they are used in photovoltaic stations.

[0003] For example, a photovoltaic panel defect detection method based on drone vision disclosed in the invention with authorization announcement number CN113284124B includes S1, drone route planning in the detection area; S2, the drone collects photovoltaic panel data to obtain infrared and visible light images; S3, infrared and visible light image registration;

[0004] S4, infrared image hot spot detection and classification; S5, defect location. This solution is not affected by the ground environment of the photovoltaic power station, and can effectively solve the problem of difficult detection of fishery-photovoltaic complementary and mountain photovoltaic power stations. It can intuitively display the defect location on the three-dimensional model, which greatly improves the maintenance efficiency of photovoltaic panels. The above technical solution has the following shortcomings: the above technical solution cannot collect large-scale orthophotos of photovoltaic panels on the planned route, and cannot array the position of photovoltaic panels, which affects the accuracy of photovoltaic panel image acquisition. At the same time, it cannot quickly identify various defects on photovoltaic panels, which is not convenient for subsequent precise maintenance of photovoltaic panels. For this reason, we propose a photovoltaic panel defect detection system and method based on drones and machine vision. Summary of the invention

[0005] The purpose of the present invention is to provide a photovoltaic panel defect detection system and method based on drones and machine vision, so as to solve the problem that the prior art proposed in the above background technology is unable to collect large-scale orthophotos of photovoltaic panels on the planned route, and is unable to perform array positioning of photovoltaic panels, which affects the accuracy of photovoltaic panel image acquisition, and is unable to quickly identify various defects on photovoltaic panels, making it inconvenient to perform subsequent precise maintenance of photovoltaic panels.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a photovoltaic panel defect detection system based on unmanned aerial vehicles and machine vision, comprising a unmanned aerial vehicle, a visible light lens and an EL lens, wherein the visible light lens or the EL lens is equipped with a gimbal and is arranged on the unmanned aerial vehicle, the unmanned aerial vehicle is connected to a control platform, the visible light lens and the EL lens are connected to an input end of an image collection module via a wireless network, the output end of the image collection module is connected to an input end of an image processing module, the output end of the image processing module is connected to an input end of a central processor, and the output end of the central processor is connected to a defect identification module;

[0007] The drone is first equipped with a visible light lens for initial high-altitude aerial photography and ground-simulating mesoscale aerial photography of automatically planned routes according to a given area; secondly, it is equipped with an EL lens for multi-view adaptive directional high-precision photography according to the route planning path;

[0008] The pan / tilt is used to adjust the EL lens at multiple angles;

[0009] The image processing module is used to perform straightening processing on the photovoltaic panel image collected by the EL lens.

[0010] Preferably, the EL lens is used to obtain high-precision images of the photovoltaic panel at most viewing angles.

[0011] Preferably, the defect identification module extracts defect features through the SimAM-EfficientNetV2 network model, classifies the sample set into defect-free samples and defective samples, and eliminates defect-free samples, and then uses the YOLOV9 network model to perform high-precision real-time detection of various defects in the image, thereby realizing rapid identification of defects.

[0012] The photovoltaic panel defect detection method based on drone and machine vision has the following specific steps:

[0013] S1. The drone is equipped with a visible light lens and automatically plans the route according to the given area through the control platform to take the first high-altitude aerial photography to obtain large-scale orthophotos and digital surface models;

[0014] S2. Use a drone equipped with a visible light lens to perform mesoscale aerial photography to obtain sub-centimeter-level precision images, identify the component area location based on the mesoscale image and the as-built drawing, and confirm the coordinates of each vertex of the component;

[0015] S3, route planning based on terrain and component coordinates, the drone is equipped with an EL lens, and uses the EL lens to perform adaptive directional shooting to obtain high-precision images of photovoltaic panels from multiple perspectives;

[0016] S4, according to the high-precision image of the multi-view photovoltaic panel, the oblique-angle image mapped in the high-precision image of the multi-view photovoltaic panel is converted into a positive-angle image through an image processing module;

[0017] S5, the central processor and the defect identification module extract defect features through the SimAM-EfficientNetV2 network model, classify the sample set into defect-free samples and defective samples, and eliminate defect-free samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid defect identification.

[0018] Preferably, in S1, the UAV automatically plans a route to perform the initial high-altitude aerial photography according to a given area range through a control platform, specifically:

[0019] According to the longitude and latitude range of the photovoltaic station or photovoltaic panel, a 300-500 meter high altitude route is planned, and a panoramic image of the photovoltaic station or photovoltaic panel and a rough digital elevation model are obtained through a visible light lens.

[0020] Preferably, in S2, the sub-centimeter precision image is obtained by performing mesoscale aerial photography of the terrain by a drone, specifically:

[0021] S20, planning a terrain-simulating mesoscale aerial photography route based on the rough digital elevation model in S1;

[0022] S21. According to the planned mesoscale aerial photography route, obtain sub-centimeter-level precision orthophotos and digital surface models through visible light lenses.

[0023] Preferably, S3 is specifically:

[0024] S30, planning a close-up shooting route for the drone according to the component area positioning, component fixed-point coordinates, and drone flight altitude information;

[0025] S31. Based on the close shooting route, multi-view high-precision images of the photovoltaic panel are collected through the EL lens.

[0026] Preferably, S4 is specifically:

[0027] Through normalization processing, the high-precision images of multi-view photovoltaic panels are converted into orthographic images and the same scale.

[0028] Preferably, S5 is specifically:

[0029] S50. Decompose the orthographic image into sample sets through the multi-scale SAM model:

[0030] The sample set is represented as:

[0031]

[0032] Where: H is the pixel height of the orthographic image; W is the pixel width of the orthographic image; h is the height of a single pixel image; w is the width of a single pixel image;

[0033] S51. Extract defect features through the SimAM-EfficientNetV2 network model, classify the sample set into non-defective samples and defective samples, and remove non-defective samples;

[0034] S52. Use the YOLOV9 network model to perform high-precision real-time detection of various defects in the image.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The present invention can collect large-scale orthophotos of photovoltaic panels on the planned route and perform array positioning of the photovoltaic panels, which greatly improves the collection accuracy of photovoltaic panel images.

[0037] (2) The present invention extracts defect features through the SimAM-EfficientNetV2 network model, classifies the sample set into non-defective samples and defective samples, and eliminates the non-defective samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid identification of defects. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 The overall schematic diagram of the photovoltaic panel defect detection system based on drone and machine vision;

[0039] Figure 2 It is a schematic diagram of the large-scale navigation of the UAV in the present invention;

[0040] Figure 3 A schematic diagram of the planned path for large-scale navigation of the UAV in the present invention;

[0041] Figure 4 It is a schematic diagram of the position structure of each vertex of the components in the geo-simulated mesoscale route of the present invention;

[0042] Figure 5 It is a sub-centimeter-level precision image map of the aerial photography of the mesoscale route of the earth in the present invention;

[0043] Figure 6 It is a schematic diagram of the process of converting the mapped oblique angle image into the orthographic angle image in the present invention;

[0044] Figure 7 is the mapped oblique-angle image collected in the present invention;

[0045] Figure 8It is the converted orthographic image in the present invention;

[0046] Fig. 9 A schematic diagram of a process for classifying a sample set into simple samples and difficult samples in the present invention;

[0047] Fig.10 It is a schematic diagram of the process of extracting defect features through the SimAM-EfficientNetV2 network model and classifying the defect features with high precision in the present invention;

[0048] Fig.11 This is an image of a photovoltaic panel after the defect sample is extracted in the present invention;

[0049] Fig.12 This is an image of a photovoltaic panel after defect samples are classified in the present invention (yellow represents hidden crack defects, and green represents black spot defects);

[0050] Fig.13 and Fig.14 It is a partial data image in the EL picture;

[0051] Fig.15 This is the accuracy curve of SimAM-EfficientNetV2;

[0052] Fig.16 Display diagram for defect prediction results;

[0053] Fig.17 This is a curve chart of YOLOV9 experimental results;

[0054] Fig.18 Display diagram for defect identification structure;

[0055] In the figure: 1. UAV; 2. gimbal; 3. EL lens; 4. control platform; 5. image collection module; 6. image processing module; 7. central processor; 8. defect identification module; 9. visible light lens. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] See also Figure 1The present invention provides a technical solution: a photovoltaic panel defect detection system based on drone and machine vision, including a drone 1, a visible light lens 9 and an EL lens 3, wherein the visible light lens 9 or the EL lens 3 is equipped with a gimbal 2 and is arranged on the drone 1, the drone 1 is connected to a control platform 4, the visible light lens 9 and the EL lens 3 are connected to an input end of an image collection module 5 through a wireless network, the output end of the image collection module 5 is connected to an input end of an image processing module 6, the output end of the image processing module 6 is connected to an input end of a central processor 7, and the output end of the central processor 7 is connected to a defect identification module 8. The visible light lens 9 in the present application uses a DJI Zenmuse P1 full-frame lens;

[0058] The drone 1 is firstly equipped with a visible light lens 9 for performing initial high-altitude aerial photography and ground-simulating mesoscale aerial photography according to the automatically planned route within a given area; secondly, it is equipped with an EL lens 3 for performing multi-view adaptive directional high-precision photography according to the route planning path;

[0059] The pan / tilt 2 is used to adjust the EL lens 3 at multiple angles;

[0060] The image processing module 6 is used to perform straightening processing on the photovoltaic panel image collected by the EL lens 3 .

[0061] Furthermore, the EL lens 3 is used to obtain multi-viewing angle high-precision images of the photovoltaic panel.

[0062] Furthermore, the defect identification module 8 extracts defect features through the SimAM-EfficientNetV2 network model, classifies the sample set into defect-free samples and defective samples, and eliminates the defect-free samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid identification of defects.

[0063] On the other hand, see Figure 2-Figure 12 , the photovoltaic panel defect detection method based on drone and machine vision, the specific steps are as follows:

[0064] Step 1, the drone 1 is equipped with a visible light lens 9 and automatically plans a route through the control platform 4 according to a given area to perform initial high-altitude aerial photography to obtain large-scale orthophotos and digital surface models;

[0065] In step 1, the drone 1 automatically plans a route to perform the first high-altitude aerial photography according to the given area range through the control platform 4, specifically:

[0066] According to the latitude and longitude range of the photovoltaic station or photovoltaic panel, a 300-500 meter high altitude flight path is planned, and a panoramic image of the photovoltaic station or photovoltaic panel and a rough digital elevation model are obtained through the visible light lens 9 for confirming the position of the array area;

[0067] Step 2: Use drone 1 to perform mesoscale aerial photography to obtain sub-centimeter-level precision images, identify the component area position based on the mesoscale image and the as-built drawing, and confirm the coordinates of each vertex of the component;

[0068] In step 2, the drone 1 is used to perform mesoscale aerial photography to obtain sub-centimeter-level precision images, specifically:

[0069] Step 20: planning a terrain-simulating mesoscale aerial photography route according to the rough digital elevation model in step 1;

[0070] Step 21: According to the planned terrain-simulating mesoscale aerial photography route, sub-centimeter-level precision orthophotos and digital surface models are obtained through the visible light lens 9 to identify the component area;

[0071] Step 3: route planning based on terrain and component coordinates, and adaptive directional shooting through EL lens 3 to obtain high-precision images of multi-view photovoltaic panels;

[0072] Step 30: Plan a close shooting route for UAV 1 according to the component area positioning, component fixed point coordinates, and UAV 1 flight altitude information;

[0073] Step 31: Based on the close-up shooting route, the EL lens 3 is used to collect multi-view high-precision images of the photovoltaic panel.

[0074] Step 4: According to the high-precision image of the multi-view photovoltaic panel, the image processing module 6 converts the oblique image in the high-precision image of the multi-view photovoltaic panel into an orthographic image;

[0075] Through normalization processing, the high-precision images of multi-view photovoltaic panels are converted into orthographic images and the same scale.

[0076] Step 5, the central processor 7 and the defect identification module 8 extract defect features through the SimAM-EfficientNetV2 network model, classify the sample set into defect-free samples and defective samples, and eliminate the defect-free samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid identification of defects.

[0077] Step 50: Decompose the orthographic image into a sample set using a multi-scale SAM model:

[0078] The sample set is represented as:

[0079]

[0080] Where: H is the pixel height of the orthographic image; W is the pixel width of the orthographic image; h is the height of a single pixel image; w is the width of a single pixel image;

[0081] Step 51, extract defect features through the SimAM-EfficientNetV2 network model, classify the sample set into simple samples and difficult samples, simple samples are defect-free samples, difficult samples are defective samples or suspected defective samples, and remove simple samples (defect-free samples);

[0082] The number of simple samples is: k;

[0083] The number of difficult samples is:

[0084] Step 52: Use the YOLOV9 network model to perform high-precision real-time detection of various defects in the image, and identify the types of defects, including single-chip short circuit, missing corners, arc cracks, through cracks, black core chips, etc.

[0085] (1) This method can collect large-scale orthophotos of photovoltaic panels on the planned route and perform array positioning of photovoltaic panels, which greatly improves the acquisition accuracy of photovoltaic panel images.

[0086] (2) This method extracts defect features through the SimAM-EfficientNetV2 network model, classifies the sample set into non-defective samples and defective samples, and eliminates non-defective samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby achieving rapid identification of defects.

[0087] (3) The drone cruises with automatic route planning, automatically adjusts the shooting angle, and automatically takes pictures.

[0088] (4) Multi-view adaptive directional high-precision shooting, introducing machine vision technology to replace visual inspection.

[0089] (5) The defect sample features are classified using the SimAM-EfficientNetV2 network model, with a defect classification accuracy of up to 90%.

[0090] (6) The YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, with a recognition accuracy of up to 88%.

[0091] The defect feature extraction process of the SimAM-EfficientNetV2 network model is given below:

[0092] 1. Take the public EL dataset as an example. The dataset contains 2,624 8-bit grayscale image samples of 300×300 pixels taken from 44 different solar cell modules with different degrees of degradation. The images in the dataset are divided into four categories, and each image is annotated with the defect probability and the type of solar module (single crystal or multi-crystalline) from which the solar cell image was originally extracted. Defects in the image are known to reduce the power efficiency of the solar module. Some EL pictures are shown in Figure 1. Fig.13 and Fig.14 shown.

[0093] Fig.13 and Fig.14 is a partial data image; Fig.13 (a), (b), (c), and (d) are single crystal silicon, and from left to right, the defect probabilities are 100%, 66.7%, 33.3%, and 0%; Fig.14 (e), (f), (g), and (h) are polysilicon. From left to right, the defect probability is 100%, 66.7%, 33.3%, and 0%.

[0094] 2. Use the EfficientNetV2 network as the backbone framework and optimize the network structure, remove the repeated layers of the Fused MBConv module and the MBConv module in the network, and reduce the number of network parameters; secondly, modify the MBConv module in the network and use large convolution kernels to replace the original convolution kernels to improve the network's receptive field; finally, use SimAM attention, which has better effects and does not increase additional parameters, to replace the SE attention module in the original network to obtain the SimAM-EfficientNetV2 network.

[0095] Table 1 shows the parameter configuration. Stage is the model stage; Operator is the module content; Stride is the stride; Channels is the number of channels of the output feature map; Layers is the number of repetitions; Conv3×3 is a regular convolution with a convolution kernel size of 3×3; the numbers after Fused MBConv and MBConv are the expansion rates; k3×3 and k5×5 are convolution kernel sizes of 3×3 and 5×5; SimAM indicates the use of the attention module; Pooling and FC are pooling and full connection.

[0096] Table 1SimAM-EfficientNetV2 network parameter configuration table

[0097]

[0098] The network as a whole does not use pre-trained weights, and adopts a non-frozen training strategy. The set training image size is 300×300. In terms of the choice of optimizer, the SGD optimizer is used, and the number of training iterations is set to 1200. The batch size is set to 32, the momentum factor is 0.9, the initial learning rate is set to 1e-2, and the weight decay is set to 5e-4. The learning rate is adjusted using the cosine annealing method. In order to prevent overfitting of the experimental data, Dropout is used during training, and its parameter is set to 0.2. The experimental accuracy is as follows Fig.15 As shown:

[0099] In order to intuitively demonstrate the classification effect of the model, the SimAM-EfficientNetV2 network model was used to perform classification prediction on the ELPV dataset. Some of the prediction results are shown in the figure below. Fig.16 Each predicted image is marked with a true label, and the corresponding prediction type and result are shown above each predicted result image, and the defect location is marked with a red circle.

[0100] The defect identification process of the YOLOV9 network model is given below:

[0101] 1. The EL dataset provided by Hebei University of Technology and Beijing University of Aeronautics and Astronautics was used to complete the training of the model. This dataset is used to benchmark abnormal defect detection methods, with a focus on industrial abnormality detection in photovoltaic cells. It contains 9 different categories of abnormal defect images, such as cracks (linear and star-shaped), broken grids, black cores, misalignment, thick lines, horizontal dislocations, lateral dislocations, printing errors, and short circuit defects.

[0102] 2. In the feature extraction part in front of the network, variable snake convolution (DSConv) is used, which will focus on optimizing the ability to extract curve features (such as cracks). DSConv can capture curve features more effectively, adapt to cracks of different shapes, and improve the accuracy and robustness of detection. Secondly, in other feature extraction parts, ScConv convolution will be used. This convolution method can significantly reduce the amount of model calculation while ensuring feature extraction capabilities, thereby improving the running speed and efficiency of the model. This improvement will help achieve faster inference time in practical applications and meet the needs of real-time detection.

[0103] The network as a whole does not use pre-trained weights, and adopts a non-frozen training strategy. The training image size is set to 640×640. In terms of the choice of optimizer, the SGD optimizer is used, and the number of training iterations is set to 400. The batch size is set to 8, the momentum factor is 0.937, the initial learning rate is set to 1e-2, and the weight decay is set to 5e-4. The learning rate is adjusted using the cosine annealing method. The experimental results are as follows Fig.17 shown.

[0104] In order to verify the effectiveness of the experimental results, the trained YOLOV9 network model is used to predict the ELPV image. Some of the prediction results are shown in the figure below. Fig.18 shown.

[0105] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. Photovoltaic panel defect detection system based on drone and machine vision, characterized by: The invention comprises an unmanned aerial vehicle (1), a visible light lens (9) and an EL lens (3), wherein the visible light lens (9) or the EL lens (3) is mounted on a pan / tilt platform (2) and arranged on the unmanned aerial vehicle (1), the unmanned aerial vehicle (1) is connected to a control platform (4), the visible light lens (9) and the EL lens (3) are connected to an input end of an image collection module (5) via a wireless network, the output end of the image collection module (5) is connected to an input end of an image processing module (6), the output end of the image processing module (6) is connected to an input end of a central processor (7), and the output end of the central processor (7) is connected to a defect identification module (8); The unmanned aerial vehicle (1) is firstly equipped with a visible light lens (9) for performing initial high-altitude aerial photography and ground-simulating mesoscale aerial photography according to an automatically planned route within a given area; secondly, it is equipped with an EL lens (3) for performing multi-view adaptive directional high-precision photography according to the route planning path; The pan / tilt platform (2) is used to adjust the EL lens (3) at multiple angles; The image processing module (6) is used to perform straightening processing on the photovoltaic panel image collected by the EL lens (3).

2. The photovoltaic panel defect detection system based on drone and machine vision according to claim 1 is characterized in that: The EL lens (3) is used to obtain multi-viewing angle high-precision images of the photovoltaic panel.

3. The photovoltaic panel defect detection system based on drone (1) and machine vision according to claim 1, characterized in that: The defect identification module (8) extracts defect features through the SimAM-EfficientNetV2 network model, classifies the sample set into defect-free samples and defective samples, and removes the defect-free samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid defect identification.

4. A photovoltaic panel defect detection method based on drone and machine vision according to any one of claims 1 to 3, characterized in that: The specific steps are as follows: S1, a drone (1) equipped with a visible light lens (9) automatically plans a route through a control platform (4) according to a given area range to perform initial high-altitude aerial photography to obtain large-scale orthophotos and digital surface models; S2, using the drone (1) to perform mesoscale aerial photography to obtain sub-centimeter-level precision images, identify the component area position based on the mesoscale image and the as-built drawing, and confirm the coordinates of each vertex of the component; S3, based on the route planning of the terrain and the component coordinates, the drone (1) is equipped with an EL lens (3), and the EL lens (3) is used to perform adaptive directional shooting to obtain high-precision images of the photovoltaic panels from multiple perspectives; S4, according to the high-precision image of the multi-view photovoltaic panel, the image processing module (6) converts the oblique-angle image mapped in the high-precision image of the multi-view photovoltaic panel into a positive-angle image; S5, through the central processor (7) and the defect identification module (8), the SimAM-EfficientNetV2 network model is used to extract defect features, classify the sample set into defect-free samples and defective samples, and eliminate the defect-free samples. Then, the YOLOV9 network model is used to perform high-precision real-time detection of various defects in the image, thereby realizing rapid defect identification.

5. The photovoltaic panel defect detection method based on drone and machine vision according to claim 4 is characterized in that: In S1, the drone (1) automatically plans a route to perform initial high-altitude aerial photography according to a given area range through the control platform (4), specifically: According to the latitude and longitude range of the photovoltaic station or photovoltaic panel, a 300-500 meter high altitude flight path is planned, and a panoramic image of the photovoltaic station or photovoltaic panel and a rough digital elevation model are obtained through a visible light lens (9).

6. The photovoltaic panel defect detection method based on drone and machine vision according to claim 4 is characterized in that: In S2, the drone (1) is used to perform mesoscale aerial photography to obtain images with sub-centimeter precision, specifically: S20, planning a terrain-simulating mesoscale aerial photography route based on the rough digital elevation model in S1; S21. According to the planned terrain-simulating mesoscale aerial photography route, sub-centimeter-level precision orthophotos and digital surface models are obtained through a visible light lens (9).

7. The photovoltaic panel defect detection method based on drone and machine vision according to claim 4 is characterized in that: The S3 is specifically: S30, planning a close shooting route for the drone (1) according to the component area positioning, the component fixed point coordinates, and the flight altitude information of the drone (1); S31, based on the close shooting route, the EL lens (3) is used to collect multi-viewing angles and high-precision images of the photovoltaic panel.

8. The photovoltaic panel defect detection method based on drone and machine vision according to claim 4, characterized in that: The S4 is specifically: Through normalization processing, the high-precision images of multi-view photovoltaic panels are converted into orthographic images and the same scale.

9. The photovoltaic panel defect detection method based on drone and machine vision according to claim 4, characterized in that: The S5 is specifically: S50. Decompose the orthographic image into sample sets through the multi-scale SAM model: The sample set is represented as: Where: H is the pixel height of the orthographic image; W is the pixel width of the orthographic image; h is the height of a single pixel image; w is the width of a single pixel image; S51. Extract defect features through the SimAM-EfficientNetV2 network model, classify the sample set into non-defective samples and defective samples, and remove non-defective samples; S52. Use the YOLOV9 network model to perform high-precision real-time detection of various defects in the image.

Images