A Photovoltaic Module EL Defect Detection Method Based on Improved YOLOv5 Network

Abstract

The present invention discloses a method for detecting EL defects of photovoltaic modules based on an improved YOLOv5 network, which establishes a detection model relying on upstream and downstream processes. Instead of taking a single cell as the detection unit, the module is used as the detection unit, and the detection and positioning of 13 common defects at present are realized. These defects not only include the defects in a single cell, but also include the overall defects with the cell module as the observation unit. Using the established photovoltaic module EL defect dataset to train and test the detection model, the experimental results show that the improved YOLOv5s cascade network model can efficiently and accurately identify various defects contained in the photovoltaic module EL image, and complete the comprehensive identification and precise positioning of complex EL defects.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic module EL defect detection methods, and specifically to a photovoltaic module EL defect detection method based on an improved YOLOv5 network. Background Art

[0002] Generally speaking, the certification work of solar photovoltaic modules consists of a series of standardized tests. Only when the module products pass all the tests can they obtain a certification certificate and be sold in the corresponding market. These tests include: appearance inspection, maximum power test, insulation withstanding voltage test, wet leakage current test, etc. More and more studies have found that the decrease in the maximum power of photovoltaic modules in the above standards is directly related to the defects found in the electroluminescence spectra of solar cells. Therefore, electroluminescence testing has been widely introduced as an important auxiliary means in current testing and certification. However, the electroluminescence testing of photovoltaic modules is different from other standard tests. Its test parameters are easy to standardize, but the test results are difficult to evaluate. Therefore, in the process of standardizing the electroluminescence testing of photovoltaic modules, the defect analysis, rating of the battery luminescence images, and the evaluation of future efficiency are both difficult points and current research hotspots. But so far, almost all photovoltaic module electroluminescence testing systems can only obtain the electroluminescence images of the modules, and it is necessary to rely on the personal experience of engineers to interpret the results. The information obtained includes: the location of the defects in the module, the types of the defects, the number of defective batteries, etc. Summary of the Invention

[0003] The purpose of the present invention is to provide a photovoltaic module EL defect detection method based on an improved YOLOv5 network, which proposes a cascaded detection network based on YOLOv5s, establishes a detection model relying on upstream and downstream processes, and uses the module rather than a single cell as the detection unit, enabling efficient and accurate detection and positioning of multiple defects simultaneously, and realizing complex EL defect recognition.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A photovoltaic module EL defect detection method based on an improved YOLOv5 network, the method comprising the following steps:

[0006] Step (1), collecting EL images of photovoltaic modules through an infrared camera, screening and sorting the image samples according to the classification standard of photovoltaic module EL defect types, and generating a sample set after manual preprocessing;

[0007] Step (2), manually annotating the sample set, converting the image data into the YOLO data set format containing target position coordinates and type information, and generating a training set;

[0008] Step (3): Import the training set generated in step (2) into the improved YOLOv5 cascade network, and perform detection training on the model to obtain the corresponding weight file.

[0009] Step (4): Use the trained model to predict the new EL image data of photovoltaic modules, mark the positions and types of defects, and conduct statistical analysis on them to determine the qualification level of the products.

[0010] Furthermore, in step (1), the dataset comes from the quality inspection pictures taken on the actual photovoltaic module production line. Since the EL images of the modules are large in area and have a lot of redundant information, it is difficult to extract small targets such as dark spots and broken grids using deep learning methods. At the same time, there is a huge difference in area between different defects. For example, the area of a dark spot and a black sheet differs by more than a hundred times, and the sample quantity is uneven. Finally, some defects can be searched with a single cell as the background, while some defects cannot be determined in a single cell. For example, a black sheet will cover the background of the cell, and the judgment of mixed bright and dark sheets needs to be based on comparison with other cells in the module. Therefore, according to the size of the area covered by different types of EL defects of photovoltaic modules, they are divided into two categories. One is that the observation background of the defect is a single cell, defined as type A defect, that is, a small target. The other is that the judgment background of the defect is based on the entire photovoltaic module, defined as type B defect, that is, a large target.

[0011] Furthermore, in step (2), for type A defects, directly mark them on the EL image of the photovoltaic module. For type B defects, first perform region segmentation and then mark them. The samples of type A and type B defects are divided into training set, test set, and validation set according to the ratio of 8:1:1 respectively. Finally, use the LabelImg tool to manually label the EL image data of the photovoltaic modules sorted out in step (1), including the type C of the EL defect and the position area coordinates of the defect, where (x1, y1) and (x2, y2) represent the upper left and lower right coordinates of the image area respectively, to obtain the corresponding YOLO format dataset as the training set available for neural network training.

[0012] Furthermore, in step (3), the YOLOv5 training network structure adopted has the following characteristics:

[0013] a. The input end adopts the Mosaic data augmentation method, which is spliced by means of random scaling, random cropping, and random arrangement to improve the detection ability for small targets. For different datasets, the anchor boxes of the initial length and width will be set. During training, the training network outputs prediction boxes based on the initial anchor boxes, and then compares them with the ground truth of the real boxes, calculates the difference between the two, and updates the network parameters in the reverse direction. That is, during each training, the optimal anchor box value in different training sets is calculated adaptively.

[0014] b. The Focus structure is designed in the reference network, and the input image is cropped through the slice operation. For example, the size of the original input image is 512*512*3. After the Slice and Concat operations, a feature map of 256*256*12 is output, and then through a Conv layer with 32 channels, a feature map of 256*256*32 is output.

[0015] Furthermore, in step (4), the specific method of the detection network is as follows: According to the size range of the areas covered by different types of photovoltaic module EL defects, they are divided into two categories. One is that the size of the defect area is within a single cell, which is defined as type A defect, that is, a small target; the other is that the defect area is distributed in an entire photovoltaic module, which is defined as type B defect, that is, a large target. According to the above definition criteria, the sample set is made into two training sets and then input into the YOLOv5 network for training respectively to obtain two weight files. When designing the detection network, we adopted a double-layer network structure, connected two detection modules in series, imported different weight files respectively to detect different targets with large differences in coverage areas, and inserted an image segmentation module between the two serially connected detection modules to improve the ability to detect targets with large differences in size simultaneously.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention adopts a photovoltaic module EL defect detection method based on YOLOv5, proposes an improved series detection network structure, realizes the recognition of complex EL defects, can detect multiple defect types, and at the same time expands the detection object from a single cell to a photovoltaic module containing multiple cells, greatly improving the detection efficiency and providing the possibility for the real-time application of the intelligent detection system in the industrial field. Description of the Drawings

[0017] Figure 1 : Flowchart of the photovoltaic module EL defect detection method based on YOLOv5 proposed by the present invention.

[0018] Figure 2 : Schematic diagram of sample segmentation.

[0019] Figure 3 : Defect classification chart.

[0020] Figure 4 : Schematic diagram of the series structure of the detection network.

[0021] Figure 5 : Detection effect display chart. Detailed Embodiments

[0022] The following will describe in detail the specific embodiments of the present invention with reference to the drawings.

[0023] As Figure 1 shown, the flowchart of the photovoltaic module EL defect detection method based on the improved YOLOv5 network includes the following steps:

[0024] Step (1): Dataset collection and processing:

[0025] Collect the EL images of photovoltaic modules through an infrared camera. According to the classification standard of photovoltaic module EL defect types, screen and sort the image samples. After manual preprocessing, a sample set is generated. The specific method is: first, conduct preliminary manual screening according to the different defect types contained in the photovoltaic module EL picture samples, and perform regional segmentation processing on the samples containing local small targets. As Figure 2 shown, it has the effect of data augmentation, and then it is sorted into a sample set.;

[0026] Step (2): Sample annotation;

[0027] For the black sheets in type A defects, directly annotate them on the EL images of photovoltaic modules, with a total of 98 samples. For type B defects, the size of the EL image of the photovoltaic module captured by the camera is 5328*3137 pixels. The EL image of the module is intercepted with sliding windows of two sizes, intercepted into two sizes of 248*505 (yellow frame) and 753*1008 (red frame) (as Figure 2 ) One EL picture of a photovoltaic module can be intercepted into 126 small images of 248*505 and 21 small images of 753*1008. Clean the images without defects among them, and the remaining ones are manually annotated. Compared with the original image samples, the network can more easily capture the defect features in the small images. Finally, 457 samples of 248*505 pixels and 492 samples of 753*1008 pixels are obtained, and the number of samples for each type of defect is not balanced. The backgrounds of these two types of samples are very different. The samples composed of six battery cells contain the gaps between the battery cells. Adding such samples can make the model have a stronger ability to distinguish the background. The samples of type A and type B defects are divided into training set, test set, and validation set according to 8:1:1 respectively. Manually annotate the sample set, convert the image data into the YOLO dataset format containing target position coordinates and type information, and generate the training set. The specific steps are: use the LabelImg tool to annotate the test set to obtain the corresponding YOLO format dataset as the training set available for neural network training.

[0028] Step (3): YOLOv5 network training model;

[0029] a. The input end adopts the method of Mosaic data augmentation, which is spliced by means of random scaling, random cropping, and random arrangement to improve the detection ability of small targets. For different data sets, the anchor boxes with initial lengths and widths will be set. During training, the training network outputs prediction boxes based on the initial anchor boxes, and then compares them with the ground truth boxes to calculate the difference between the two, and then updates them in reverse to iterate the network parameters. That is, during each training, the optimal anchor box values in different training sets are adaptively calculated.

[0030] b. A Focus structure is designed in the benchmark network, and the input image is cropped through the slice operation. For example, the size of the original input image is 512*512*3. After the Slice and Concat operations, a feature map of 256*256*12 is output, and then through a Conv layer with 32 channels, a feature map of 256*256*32 is output.

[0031] Step (4): Detection model;

[0032] Use the trained model to predict the new photovoltaic module EL image data, mark the location and type of defects, and conduct statistical analysis on them to judge the qualification level of the product. The specific method of the detection network is as follows: According to the size range of the areas covered by different types of photovoltaic module EL defects, they are divided into two categories. One is that the size area of the defect is within a single cell, which is defined as type A defect, that is, a small target; the other is that the defect area is distributed in an entire photovoltaic module, which is defined as type B defect, that is, a large target. According to the above definition criteria, the sample set is made into two training sets and then input into the YOLOv5 network for training respectively to obtain two weight files. When designing the detection network, we adopted a double-layer network structure, connected two detection modules in series, imported different weight files respectively to detect different targets with large differences in coverage areas, and inserted an image segmentation module between the two series-connected detection modules to improve the ability to detect targets with large differences in size at the same time. The detection process is as Figure 4 shown. The image to be detected is a complete photovoltaic module EL image. After being input into the detection system, it first passes through the first-layer detection module to mark type A defects, then passes through segmentation, and then is passed into the second-layer detection module to mark type B defects. Finally, it is restored to output the complete detection result. To sum up, in order to achieve the accuracy and real-time performance of detection and liberate manual labor from the cumbersome detection work, a series-connected detection model based on the Yolov5s network is proposed. Through experimental testing, the accuracy, speed, and categories of this method meet the requirements of industrial production, especially meeting the current status of component detection in the quality inspection link, which proves that the method of the present invention has certain advantages.

[0033] Finally, it should be noted that the above embodiments are only a detailed clarification of the technical solution, rather than a limitation thereof; although the technical solution of the present invention has been described with reference to the specific embodiments, those of ordinary skill in the art should understand; they can still modify the solutions of the embodiments, or make equivalent substitutions for some of them; such modifications or substitutions do not cause their essence to deviate from the scope of the technical solution proposed by the present invention, and should all be included in the scope of the claims and the specification of the present invention.

Claims

1. A photovoltaic module EL defect detection method based on an improved YOLOv5 network, characterized in that It includes the following steps: Step (1): Collect the EL images of photovoltaic modules through an infrared camera. According to the classification standard of EL defect types of photovoltaic modules, screen and sort the image samples. After manual preprocessing, a sample set is generated; Step (2): Manually annotate the sample set, convert the image data into the YOLO dataset format containing target position coordinates and type information, and generate a training set; Step (3): Import the training set generated in step (2) into the improved YOLOv5 cascade network, train the model, and obtain the corresponding weight file; Step (4): Use the trained model to predict the new EL image data of photovoltaic modules, mark the positions and types of defects, and conduct statistical analysis on them to judge the qualified level of the products; In step (1), according to the area range covered by different types of EL defects of photovoltaic modules, they are divided into two categories. One is that the observation background of the defect is a single cell, which is defined as type A defect, that is, a small target. The other is that the judgment background of the defect is based on the entire photovoltaic module, which is defined as type B defect, that is, a large target; In step (2), for type A defects, directly mark them on the EL image of the photovoltaic module; for type B defects, first perform region segmentation and then mark them; divide the samples of type A and type B defects into training set, test set, and validation set according to 8:1:1 respectively; finally, use the LabelImg tool to manually annotate the test set, including the EL defect type C and the position area coordinates of the defect, where (x1, y1) and (x2, y2) represent the upper left corner and the lower right corner coordinates of the image area respectively, and obtain the corresponding YOLO format dataset as the training set available for neural network training; In step (4), when designing the detection network, a two-layer network structure is adopted, two detection modules are cascaded, different weight files are imported respectively to detect different targets with large differences in coverage areas, and an image segmentation module is inserted between the two cascaded detection modules to improve the ability to detect targets with large size differences simultaneously; 2. The photovoltaic module EL defect detection method based on the improved YOLOv5 network according to claim 1, wherein: In step (3), the input end of the YOLOv5 training network adopts the Mosaic data augmentation method, which is spliced by means of random scaling, random cropping, and random arrangement. For different datasets, the initial anchor boxes of length and width are set; during training, the training network outputs prediction boxes based on the initial anchor boxes, and then compares them with the true boxes groundtruth, calculates the difference between the two, and updates them backward to iterate the network parameters, that is, each time during training, the optimal anchor box values in different training sets are calculated adaptively; 3. The photovoltaic module EL defect detection method based on the improved YOLOv5 network according to claim 1, characterized in that: In step (3), in the YLOLv5 network, a Focus structure is designed to crop the input image through slice operation.

Images