Photovoltaic module hot spot inspection method based on unmanned aerial vehicle and infrared image processing

Abstract

The invention discloses a photovoltaic module hot spot inspection method based on an unmanned aerial vehicle and infrared image processing. The method comprises the following steps: S1, shooting an infrared image of a photovoltaic module of a photovoltaic power station through the unmanned aerial vehicle; s2, the acquired infrared image and the acquired GPS information are processed; and S3, generating a fault diagnosis result. Wherein S2 comprises the following steps: removing an interference background of the infrared image by adopting an improved DeepLabv3 + model, and extracting a photovoltaic module area; and carrying out hot spot fault detection on the infrared image after the background is removed by adopting an improved YOLOv5 algorithm, and accurately identifying hot spots. According to the two-stage hot spot detection method based on deep learning, the influence of a complex background in the infrared image of the photovoltaic module on hot spot detection can be effectively eliminated, and the YOLOv5 target detection model is improved, so that the performance of small target hot spot detection is improved, and accurate identification of hot spots is realized.

Description

Technical Field

[0001] The present invention relates to the photovoltaic field, and more specifically to a photovoltaic component hot spot inspection method based on unmanned aerial vehicle and infrared image processing. Background Art

[0002] In order to cope with climate change, various countries have vigorously developed renewable energy, among which photovoltaics, as an important part of renewable energy, has received special attention. According to the data of the "China Renewable Energy Development Report 2022", by the end of 2022, the cumulative installed capacity of photovoltaics will increase by 28.6% year-on-year, and the national solar power generation will increase by 30.8% year-on-year, accounting for 4.9% of the total annual power generation of all power sources. Photovoltaics are mostly in natural environments, and foreign objects such as dust, fallen leaves, and bird feces will inevitably appear on the surface. If they are not cleaned in time, the shaded battery cells will act as loads, absorb current and convert it into heat to form hot spots. When the temperature reaches a certain level, a series of disasters such as fire and explosion may occur. Therefore, it is very important to propose an efficient photovoltaic hot spot detection strategy and timely detect and locate hot spot failures in photovoltaic cell modules for the daily maintenance of photovoltaic power stations.

[0003] Currently, the commonly used hot spot detection methods can be divided into detection methods based on photovoltaic electrical characteristics and fault detection methods based on photovoltaic infrared images. The hot spot detection method based on photovoltaic electrical characteristics requires the construction of peripheral circuits and the use of sensors for data monitoring, and uses the changes in photovoltaic output characteristics under fault conditions to detect hot spot faults. The implementation principle of this method is simple, but for large-scale photovoltaic power stations, thousands of sensors are required, which is costly and inefficient. At the same time, the detection results of this method are greatly affected by the weather, and the specific hot spot position cannot be accurately located.

[0004] The fault detection method based on photovoltaic infrared images uses infrared thermal imagers to detect the temperature distribution on the surface of photovoltaic modules in order to identify potential faults or abnormal conditions, as the temperature of photovoltaic modules varies under different working conditions. This method does not require contact with photovoltaic modules, is low-cost, and is significantly more efficient than the detection method based on electrical characteristics.

[0005] With the development of deep learning related technologies and drone hardware, it is possible to use drones to carry infrared imagers for automatic inspection, collect infrared images, and then solve the hot spot problem through image processing. Guo Menghao improved the Faster RCNN model by combining transfer learning, and highlighted the hot spot area through image preprocessing, further improving the detection accuracy. Sun Hairong used a deep transfer learning model to solve the problem of the scarcity of hot spot samples. IMENES AG synthesized visible light and infrared images and used composite images to train the YOLO network. This method can successfully detect and locate weak fault features, but the fault classification effect is poor. HONG F combined the use of visible light images for segmentation and infrared images for fault diagnosis. Wang Daolei used style transfer to expand the infrared hot spot image dataset and applied the multi-scale feature learning module SpotFPN to Faster RCNN to further improve the detection accuracy. Wang Daolei reconverted color images into grayscale images to expand the dataset. At the same time, a new attention network was proposed, and the detection accuracy of the improved SSD model was further improved.

[0006] Although existing deep learning methods have achieved good results in detecting hot spot faults, there are still several problems that need to be solved: (1) The infrared characteristics of the interference background and hot spot faults in aerial photovoltaic infrared images are very similar; (2) Photovoltaic hot spot faults are mostly small targets with low detection accuracy; (3) There is currently no open source photovoltaic hot spot fault dataset.

[0007] Therefore, there is an urgent need in the art for a method for detecting hot spot failures of photovoltaic modules that can effectively solve the above-mentioned problems. Summary of the invention

[0008] The purpose of the present invention is to provide a hot spot inspection method for photovoltaic modules based on drones and infrared image processing. The present invention can effectively eliminate the influence of complex background in infrared images of photovoltaic modules on hot spot detection through a two-stage hot spot detection method based on deep learning, and improve the performance of small target hot spot detection by improving the YOLOv5 target detection model, thereby achieving accurate identification of hot spots.

[0009] The present invention provides a method for hot spot inspection of photovoltaic modules based on unmanned aerial vehicles and infrared image processing, the method comprising:

[0010] S1. Taking an infrared image of the photovoltaic components of the photovoltaic power station by a drone;

[0011] S2. Processing the collected infrared image and GPS information; and

[0012] S3. Generate fault diagnosis results;

[0013] S2 includes:

[0014] An improved DeepLabv3+ model is used to remove the interfering background of the infrared image and extract the photovoltaic module area;

[0015] The improved YOLOv5 algorithm is used to detect hot spot faults on the infrared image after the background is removed to accurately identify hot spots.

[0016] In another preferred example, the infrared image of the photovoltaic components of the photovoltaic power station is captured and collected by an infrared imager carried by the drone.

[0017] In another preferred example, the inspection route of the drone is planned according to the map image of the photovoltaic power station.

[0018] In another preferred embodiment, the drone flies autonomously along a preset inspection route, and sets the shooting height and angle to avoid various interferences such as reflections and changes in the size of photovoltaic modules.

[0019] In another preferred example, the drone collects infrared images in real time.

[0020] In another preferred embodiment, the drone transmits the captured images and GPS information to a ground station via a wireless network.

[0021] In another preferred embodiment, the ground station processes the infrared image and GPS information of the photovoltaic module after receiving them.

[0022] In another preferred embodiment, S2 includes:

[0023] S21. Producing a data set of the infrared image of the photovoltaic module;

[0024] S22. Establish a photovoltaic module region segmentation model and train it so that the infrared image processed by S21 is input and the photovoltaic module region segmentation result is output; mask processing is performed on the area outside the photovoltaic module to obtain a mask image that does not contain a natural background;

[0025] S23. Establish a hot spot fault detection model and train it so that the mask image that does not contain the natural background is input and the hot spot area positioning result is output.

[0026] In another preferred example, S21 further includes: rotating, flipping, translating, changing brightness and adding Gaussian noise to the original infrared image to expand the data set.

[0027] In another preferred example, S22 also includes: using LabelMe software to mark the pixels of the photovoltaic component area of ​​the infrared image to produce a semantic segmentation data set, wherein the photovoltaic component area is a positive sample and the remaining areas are negative samples; using the segmentation data set to train the improved DeepLabv3+ model to establish a photovoltaic component area segmentation model.

[0028] In another preferred example, S23 also includes: screening out infrared images with hot spot faults from the mask image that does not contain the background area, marking the hot spot fault category using LabelImg software to produce a target detection data set; using the target detection data set to train the improved YOLOv5 model to establish a hot spot fault detection model.

[0029] In another preferred embodiment, the improved DeepLabv3+ model includes: replacing the backbone network Xception of DeepLabv3+ with the MobileNetv2 network; and changing the downsampling multiple of the encoder to the input image from 16 to 8; replacing its cross entropy loss function with a hybrid loss function L consisting of a cross entropy loss function and a Dice Loss loss function. seg :

[0030]

[0031] Where: L CE ——Cross entropy loss function; L Dice ——Dice Loss loss function; r——weight factor, set to 0.5.

[0032] In another preferred embodiment, the cross entropy loss function L CE for:

[0033] L CE =-[ylog(p)+(1-y)log(1-p)] (1)

[0034] Where: y——label; p——the probability that a sample is predicted to be a positive sample.

[0035] In another preferred embodiment, the Dice Loss function L Dice for:

[0036]

[0037] Where: y∩p is the number of elements shared by y and p.

[0038] In another preferred example, the input image is the original infrared image.

[0039] In another preferred example, the improved YOLOv5 model includes: adopting an SPPCSPC module; introducing a shallow detection scale and recalculating the size of the anchor box using a K-Means clustering algorithm; and improving a border regression loss function to measure the similarity between a predicted border and a real border.

[0040] In another preferred example, a shallow detection scale is introduced, and the size of the anchor box is recalculated using the K-Means clustering algorithm, including introducing a small target detection layer, and the shallow feature map is spliced ​​with the deep feature map for detection, which is expanded to four-scale detection.

[0041] In another preferred example, the shallow information of the small target detection layer includes a lot of information such as shape, position and size.

[0042] In another preferred embodiment, the frame regression loss function L is

[0043] L=L obj +L cls +L box ;

[0044] Where: L obj ——Confidence loss, using cross entropy loss; L cls ——Classification loss, using cross entropy loss; L box ——Border regression loss, using SIoU Loss.

[0045] In another preferred embodiment,

[0046] Where: L IoU ——IoU loss function; Δ——distance loss; Ω——shape loss.

[0047] In another preferred example, the fault diagnosis result includes a fault diagnosis report and a location mark of the faulty photovoltaic component.

[0048] In another preferred embodiment, a fault diagnosis report is generated for each inspection result.

[0049] In another preferred embodiment, the GPS information of the faulty photovoltaic module is displayed on a visual interface to achieve marking of the faulty photovoltaic module.

[0050] In another preferred embodiment, the fault information is pushed to the staff's mobile phone APP to remind maintenance.

[0051] The main advantages of the present invention include:

[0052] The present invention adopts a two-stage hot spot detection method based on deep learning, which can effectively eliminate the influence of complex background in infrared images of photovoltaic modules on hot spot detection. By improving the YOLOv5 target detection model, the performance of small target hot spot detection is improved, and accurate identification of hot spots can be achieved.

[0053] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0055] Figure 1 is a schematic diagram of a photovoltaic module hot spot inspection system in an example of the present invention;

[0056] Figure 2 It is an overall framework diagram of a photovoltaic module infrared image hot spot detection method in an example of the present invention;

[0057] Figure 3 is a DeepLabv3+ model structure diagram in an example of the present invention;

[0058] Figure 4 It is a structural diagram of an improved YOLOv5 model in an example of the present invention;

[0059] Figure 5 It is a structural diagram of the SPPCSPC model in an example of the present invention;

[0060] Figure 6 (a) shows the detection effect of the original YOLOv5 model without removing the background.

[0061] Figure 6 (b) is a diagram showing the detection effect of the two-stage hot spot detection method of the present invention. DETAILED DESCRIPTION

[0062] After extensive and in-depth research and a large number of screenings, the inventors have developed for the first time a hot spot inspection method for photovoltaic components based on drones and infrared image processing. Specifically, this application uses a drone equipped with an infrared imager to collect infrared images of a photovoltaic power station to create a self-made hot spot fault data set, and proposes a two-stage hot spot detection method based on deep learning; the first stage removes interference background, and proposes an improved DeepLabv3+ segmentation model by replacing the backbone network and using a hybrid loss function to address the problems of slow segmentation speed, poor edge extraction effect, and imbalance of positive and negative samples of photovoltaic components, so as to achieve fast and accurate extraction of photovoltaic component areas; the second stage performs hot spot fault detection, and proposes an improved YOLOv5 hot spot detection model by using an enhanced SPP module, introducing a shallow detection scale, and changing the border regression loss function to address the problems of missed detection and false detection of small target hot spots, thereby achieving accurate identification of hot spots. The present invention was completed on this basis.

[0063] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, the accompanying drawings are schematic diagrams, so the device and apparatus of the present invention are not limited by the size or ratio of the schematic diagrams.

[0064] It should be noted that in the claims and description of this patent, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one" do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0065] Example

[0066] Hot spot faults in aerial photovoltaic infrared images are mostly small targets and are very similar to the interference background, resulting in low hot spot fault detection accuracy. To overcome this problem, this embodiment proposes a photovoltaic module hot spot inspection method based on drones and infrared image processing.

[0067] The photovoltaic module hot spot inspection method comprises:

[0068] (1) The inspection route is planned according to the map image of the photovoltaic power station. The drone flies autonomously according to the preset route, and sets the shooting height and angle to avoid various interferences such as reflection and changes in the size of photovoltaic modules;

[0069] (2) The drone collects infrared images in real time and transmits the captured images and GPS information to the ground station via a wireless network;

[0070] (3) After the ground station receives the infrared image of the photovoltaic module, it first uses the improved DeepLabv3+ algorithm to remove the interference background and extract the photovoltaic module area. Then, it uses the improved YOLOv5 algorithm to detect hot spot faults on the infrared image after removing the background and accurately identify the hot spot.

[0071] (4) A fault diagnosis report is generated for each inspection result, and the GPS information of the fault point is displayed on a visual interface to mark the faulty PV module. At the same time, the system pushes the fault information to the staff’s mobile phone APP to remind them to perform maintenance.

[0072] Among them, the GPS information recorded when encountering a fault is the latitude and longitude coordinates and altitude information of the current position of the drone when taking pictures. However, combined with the GPS information and attitude data of the drone (such as pitch angle, yaw angle, etc.), the fault point location information in the image can be converted into the spatial coordinate information of the actual photovoltaic module fault point. Therefore, although the GPS information recorded by the drone directly reflects the location information of the drone when shooting, the precise location information of the photovoltaic module fault point can be indirectly obtained through a series of technical means.

[0073] This method is implemented based on the photovoltaic module hot spot inspection system using drones and infrared image processing. Figure 1 As shown in the figure, the system includes a UAV and a ground station that communicate with each other. The UAV is used to collect infrared images in real time and transmit the captured images and GPS information to the ground station via a wireless network. After the ground station receives the infrared image of the photovoltaic module, it uses a two-stage method based on deep learning to detect hot spots:

[0074] In the first stage, interference background removal is performed. To address the problems of slow segmentation speed, poor edge extraction effect, and imbalanced positive and negative samples of photovoltaic modules, an improved DeepLabv3+ segmentation model is proposed by replacing the backbone network and adopting a hybrid loss function to achieve fast and accurate extraction of photovoltaic module areas. In the second stage, hot spot fault detection is performed. To address the problems of missed detection and false detection of hot spots of small targets, an improved YOLOv5 hot spot detection model is proposed by adopting an enhanced SPP module, introducing a shallow detection scale, and changing the border regression loss function to achieve accurate identification of hot spots.

[0075] The overall framework of the two-stage photovoltaic infrared image hot spot detection method proposed in this invention is as follows: Figure 2 As shown:

[0076] 1) First, a dataset is created by collecting infrared images of photovoltaic modules using an infrared imager mounted on a drone. The original images are randomly rotated, flipped, translated, brightness changed, and Gaussian noise added to expand the dataset.

[0077] 2) Secondly, establish a component segmentation model. Use LabelMe software to mark the pixels of the photovoltaic component area to produce a semantic segmentation dataset, where the photovoltaic component area is a positive sample and the rest of the area is a negative sample. Use the segmentation dataset to train the improved DeepLabv3+ model and establish a photovoltaic component area segmentation model. After the training is completed, input the original image, output the photovoltaic component area segmentation result, and mask the area outside the photovoltaic component to obtain a mask image that does not contain the natural background.

[0078] 3) Finally, a hot spot fault detection model is established. Infrared images with hot spot faults are screened out from the mask images that do not contain the background area obtained in the first stage, and the hot spot fault categories are marked using LabelImg software to create a target detection dataset. The improved YOLOv5 model is trained using the target detection dataset to establish a hot spot fault detection model. After the training is completed, a mask image that does not contain the natural background is input to obtain the output hot spot area positioning result.

[0079] The two-stage hot spot detection method based on deep learning can effectively eliminate the influence of complex background in the infrared image of photovoltaic modules on hot spot detection. By improving the YOLOv5 target detection model, the performance of small target hot spot detection is improved, and accurate identification of hot spots can be achieved.

[0080] 1. Photovoltaic module area segmentation model

[0081] The structure diagram of the DeepLabv3+ model is as follows Figure 3 As shown in the figure, the original DeepLabv3+ model extracts photovoltaic modules slowly and the edges of the photovoltaic modules are not smooth. Since the speed and accuracy of extracting the photovoltaic module area affect the hot spot fault detection in the second stage, considering the high dependence of the two-stage hot spot detection method on the segmentation results of the first stage, the present invention replaces the backbone network Xception of DeepLabv3+ with the MobileNetv2 network to reduce the number of model parameters and speed up the segmentation speed; the encoder downsampling multiple of the input image is changed from 16 to 8 to reduce information loss, so that the network can better retain the high-resolution information of the input image and improve the segmentation effect.

[0082] Photovoltaic modules occupy a relatively small proportion of pixels in infrared images, which will lead to the problem of imbalance between positive and negative samples. The original cross entropy loss function often cannot effectively balance the categories of fewer samples. The cross entropy loss function L CE The formula is shown in (1). Dice Loss is often used in the case of imbalanced positive and negative samples. It can effectively handle the difference in the number of pixels between the target area and the background. The Dice Loss loss function L Dice The formula is shown in (2). In order to solve the problem of imbalance between positive and negative samples and improve segmentation accuracy, the segmentation model of the present invention uses a hybrid loss function L consisting of a cross entropy loss and a Dice Loss function. seg , the mixed loss function L seg The formula is shown in (3).

[0083] L CE =-[ylog(p)+(1-y)log(1-p)] (3)

[0084] Where: y——label; p——the probability that a sample is predicted to be a positive sample.

[0085]

[0086] Where: y∩p is the number of elements shared by y and p.

[0087]

[0088] Where: r is the weight factor of the loss function, set to 0.5.

[0089] 2. Photovoltaic module hot spot detection model

[0090] In order to solve the problem of false detection and missed detection of small target hot spots in aerial photovoltaic infrared images, this paper adopts an enhanced version of the SPP module based on the YOLO5s model to expand the receptive field and enhance feature extraction and expression; introduces a shallow detection scale and uses the K-Means clustering algorithm to recalculate the size of the anchor box to increase multi-scale perception capabilities and improve the detection performance of small target hot spots; improves the bounding box regression loss function to more accurately measure the similarity between the predicted bounding box and the real bounding box. The specific model structure diagram after improvement is shown in the figure below. Figure 4 shown.

[0091] 1) Use enhanced SPP module: SPPCSPC

[0092] Photovoltaic infrared images have low resolution and cannot provide subtle details and accurate spatial information. In addition, the color contrast between the hot spot area and the photovoltaic panel is low, and there is a problem of noise interference. Therefore, this paper uses the SPPCSPC module of YOLOv7 to replace the original SPP structure of YOLOv5s in the Backbone part. The optimized model expands the receptive field through pooling operations, enriches the features of the missing areas, and enhances the model's feature fusion ability for images with different degrees of integrity of the effective feature areas. Figure 5 shown.

[0093] 2) Introducing a shallow detection layer

[0094] The YOLOv5s model uses three-scale feature layers for detection, with scales of 20×20, 40×40, and 80×80. Since the downsampling multiple of the YOLOv5 model is relatively large, the deeper feature maps will lose a lot of feature information of small targets after multiple convolutions and pooling during the training process, resulting in reduced detection accuracy.

[0095] In order to solve the above problems, the present invention introduces a small target detection layer, which performs detection after splicing the shallow feature map with the deep feature map, and expands it to four-scale detection. The improved structure is as follows Figure 3 As shown in the figure. After the 17th layer, convolutional layers and upsampling are continued to be added to expand the size of the feature map. At the same time, at the 20th layer, the feature map of size 160×160 is concatenated with the feature map of the second layer in the backbone network to obtain richer and multi-scale feature maps for small target detection. The introduced shallow information contains a lot of information such as shape, position and size, which reduces the information loss of the features of the smaller target area during the convolution process and can better detect small targets.

[0096] 3) Improve the loss function

[0097] The loss of YOLOv5 mainly consists of three parts, and the calculation formula is shown in (4):

[0098] L=L obj +L cls +L box (6)

[0099] Where: L obj ——Confidence loss, using cross entropy loss; L cls ——Classification loss, using cross entropy loss; L box ——Border regression loss, using CIoU Loss.

[0100] CIoU Loss considers the overlapping area, center point distance and aspect ratio of bounding box regression, but the aspect ratio describes a relative value, which is somewhat vague and does not consider the balance between difficult and easy samples. Therefore, the present invention selects SIoU instead of CIoU as the regression loss of the prediction box.

[0101] SIoU consists of four parts: angle cost, distance cost, shape cost, and IoU cost. The formula is:

[0102]

[0103] Where: L IoU ——IoU loss function; Δ——distance loss; Ω——shape loss.

[0104] Compared with other IoU loss functions, the SIoU loss function not only considers distance, shape and IoU in bounding box regression, but also introduces the vector angle required for regression, that is, the matching direction, redefines the distance loss, and effectively reduces the degrees of freedom of regression, thereby accelerating the convergence of the network.

[0105] In order to further verify the detection performance of the two-stage aerial infrared image hot spot detection method, this paper selects pictures from the test set for testing. The test results are as follows: Figure 6 As shown, Figure 6 (a) shows the detection effect of the original YOLOv5 model without removing the background. Figure 6 (b) is the detection effect diagram using the two-stage hot spot detection method. It can be clearly seen that the infrared characteristics of the interference background are very similar to hot spot faults. The original YOLOv5 model will mistake the interference background for hot spot faults, but the two-stage hot spot detection method will first quickly and accurately extract the photovoltaic module area from the infrared image to remove the impact of background interference on hot spot detection. In addition, the improved YOLOv5 model can detect faults that the original YOLOv5 did not detect, and the overall confidence of hot spot fault detection is higher than that of the original YOLOv5 model.

[0106] All documents mentioned in the present invention are cited as references in this application, just as each document is cited as reference individually. In addition, it should be understood that after reading the above teachings of the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms also fall within the scope defined by the claims attached to this application.

Claims

1. A method for hot spot inspection of photovoltaic modules based on drones and infrared image processing, characterized in that: The method comprises: S1. Taking an infrared image of the photovoltaic components of the photovoltaic power station by a drone; S2. Processing the collected infrared image and GPS information; and S3. Generate fault diagnosis results; S2 includes: An improved DeepLabv3+ model is used to remove the interfering background of the infrared image and extract the photovoltaic module area; The improved YOLOv5 algorithm is used to detect hot spot faults on the infrared image after the background is removed to accurately identify hot spots.

2. The method according to claim 1, characterized in that S2 include: S21. Producing a data set of the infrared image of the photovoltaic module; S22. Establish a photovoltaic module region segmentation model and train it so that the infrared image processed by S21 is input and the photovoltaic module region segmentation result is output; mask processing is performed on the area outside the photovoltaic module to obtain a mask image that does not contain a natural background; S23. Establish a hot spot fault detection model and train it so that the mask image that does not contain the natural background is input and the hot spot area positioning result is output.

3. The method according to claim 2, characterized in that S21 also includes: rotating, flipping, translating, changing brightness and adding Gaussian noise to the original infrared image to expand the data set.

4. The method according to claim 2, characterized in that S22 also includes: using LabelMe software to mark the pixels of the photovoltaic component area of ​​the infrared image to produce a semantic segmentation data set, wherein the photovoltaic component area is a positive sample and the remaining areas are negative samples; using the segmentation data set to train the improved DeepLabv3+ model to establish a photovoltaic component area segmentation model.

5. The method according to claim 2, characterized in that S23 also includes: screening out infrared images with hot spot faults from the mask image that does not contain the background area, marking the hot spot fault category using LabelImg software to produce a target detection data set; using the target detection data set to train the improved YOLOv5 model to establish a hot spot fault detection model.

6. The method according to claim 1, characterized in that The improved DeepLabv3+ model includes: replacing the backbone network Xception of DeepLabv3+ with the MobileNetv2 network; changing the downsampling multiple of the encoder for the input image from 16 to 8; replacing its cross entropy loss function with a hybrid loss function L consisting of a cross entropy loss function and a Dice Loss loss function. seg : Where: L CE ——Cross entropy loss function; L Dice ——Dice Loss loss function; r——weight factor, set to 0.

5.

7. The method according to claim 1, characterized in that The improved YOLOv5 model includes: adopting the SPPCSPC module; introducing a shallow detection scale and recalculating the size of the anchor box using the K-Means clustering algorithm; and improving the border regression loss function to measure the similarity between the predicted border and the real border.

8. The method according to claim 7, characterized in that A shallow detection scale is introduced, and the K-Means clustering algorithm is used to recalculate the size of the anchor box, including introducing a small target detection layer, splicing the shallow feature map with the deep feature map for detection, and expanding it to four-scale detection.

9. The method according to claim 7, characterized in that The border regression loss function L is L=L obj +L cls +L box ; Where: L obj ——Confidence loss, using cross entropy loss; L cls ——Classification loss, using cross entropy loss; L box ——Border regression loss, using SIoU Loss.

10. The method according to claim 9, characterized in that Where: L IoU ——IoU loss function; Δ——distance loss; Ω——shape loss.

Images