Photovoltaic cell defect detection method and device, computer equipment and storage medium

Abstract

The embodiment of the invention discloses a photovoltaic cell defect detection method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring an EL image of a photovoltaic cell to be detected; inputting the image into a photovoltaic cell defect detection model for defect detection and classification; wherein the photovoltaic cell defect detection model is a model obtained by acquiring an original photovoltaic cell EL image, expanding the original photovoltaic cell EL image by using a StyleGAN3-T generative adversarial network, performing preprocessing, labeling and data enhancement on the expanded photovoltaic cell EL image to obtain a sample set, and training a deep learning network; and outputting a detection result. By implementing the method provided by the embodiment of the invention, the problems that the existing photovoltaic monitoring technology mainly depends on the traditional image processing method, the efficiency is low, errors are easy to make and the expansion is difficult can be solved, and the challenges that the data set is small, the universality is poor and the performance is insufficient in a complex and changeable environment can be solved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic monitoring and defect detection, and in particular to a photovoltaic cell defect detection method, device, computer equipment and storage medium. Background Art

[0002] In recent years, the installation and application of photovoltaic systems have shown a rapid growth trend. Although solar technology, especially large-scale photovoltaic systems, has made significant progress, some problems still exist, such as performance degradation and poor reliability. These problems are usually caused by production defects, improper operation and extreme weather conditions, resulting in equipment performance degradation, early failure and even fire risks. In particular, silicon cells in photovoltaic modules are prone to cracking or other defects during the production stage, which will gradually affect the overall performance of the system over time. Even minor manufacturing defects may cause serious degradation during use. In addition, as photovoltaic systems are exposed to extreme environmental conditions (such as heat load, mechanical stress, humidity, salt spray, etc.), other types of damage may occur, further exacerbating the degradation of components. Due to these problems, the unstable and unreliable performance of photovoltaic systems has brought many challenges to the industry and hindered the sustainable development of energy and the environment.

[0003] To address these issues, photovoltaic systems can be monitored at different life cycle stages. Currently, widely used monitoring methods include electroluminescence, infrared thermal imaging, photoluminescence, visible light detection, spectroscopy, and current-voltage characteristic curves. Among these methods, electroluminescence imaging technology is one of the most commonly used indoor monitoring methods. However, these imaging technologies usually rely on traditional or manual methods, and have many problems such as limited application scale, prone to errors, low efficiency, and high cost. Traditional manual methods are gradually becoming incapable of handling large numbers of images and requiring high levels of professional visual analysis. For this reason, automated photovoltaic monitoring technology has been proposed as an accurate and efficient solution with better economy.

[0004] In recent years, automated monitoring techniques for different imaging methods have been widely studied. However, due to many limitations in practical applications, these technologies have not yet been popularized on a commercial and industrial scale and are still in the laboratory research stage. Early computer vision methods mainly relied on manual feature extraction and image processing to detect and locate defects. These methods face some challenges in practical applications, such as insufficient performance, poor versatility, lack of end-to-end operation process, and the need for manual intervention and additional steps. In addition, existing methods usually rely on small data sets, have limited defect coverage, and have difficulty in coping with various challenges in real-time applications.

[0005] With the introduction of machine learning and deep learning techniques, the application of computer vision in photovoltaic monitoring has been significantly improved. Compared with traditional methods, machine learning and deep learning can provide more robust and reliable solutions. However, despite the achievements in some small-scale applications, existing methods still face a series of problems in commercial and industrial applications. For example, how to handle smaller data sets, expand the scope of application, cope with detection tasks under unknown working conditions, identify complex crack types in low-contrast backgrounds, improve performance and robustness, accurately locate and detect defects, avoid missed detection, and cope with the challenges of multiple defects. Ensuring the applicability and resilience of the detection framework in a changing environment and its effectiveness against different types of defects will be the key to achieving large-scale applications.

[0006] Therefore, it is necessary to design a new method to solve the problems that existing photovoltaic monitoring technology mainly relies on traditional image processing methods, which are inefficient, error-prone and difficult to expand, and even machine learning and deep learning technologies still face the challenges of small data sets, poor versatility and insufficient performance in complex and changing environments. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the prior art and provide a photovoltaic cell defect detection method, device, computer equipment and storage medium.

[0008] To achieve the above object, the present invention adopts the following technical solution: a photovoltaic cell defect detection method, comprising: Acquire the EL image of the photovoltaic cell to be inspected; The photovoltaic cell EL image to be detected is input into the photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is a model obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network; The detection result is outputted.

[0009] The further technical solution is: the photovoltaic cell defect detection model is obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network to obtain the model, including: Collect original photovoltaic cell EL images; Using StyleGAN3-T to generate adversarial networks to expand the original photovoltaic cell EL image to obtain an expanded photovoltaic cell EL image; Preprocessing and marking the expanded photovoltaic cell EL image to obtain a marking result; Performing data enhancement on the labeling result to obtain a sample set; Build deep learning networks; The deep learning network is trained using the sample set to obtain a photovoltaic cell defect detection model.

[0010] The further technical solution is: the StyleGAN3-T generative adversarial network is used to expand the original photovoltaic cell EL image to obtain an expanded photovoltaic cell EL image, including: Set the seed to initialize the random generator of the StyleGAN3-T generative adversarial network and set the set cutoff value; Based on the original photovoltaic cell EL image, a StyleGAN3-T generative adversarial network is used to generate a synthetic image to obtain an expanded photovoltaic cell EL image; Among them, the StyleGAN3-T generative adversarial network is a pre-trained model that is trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

[0011] A further technical solution is: performing data enhancement on the labeling result to obtain a sample set includes: Mosaic is used to perform data enhancement on the labeling results to obtain a sample set.

[0012] Its further technical solution is: the deep learning network includes a Yolov9 model with programmable gradient information and a generalized efficient layer aggregation network structure.

[0013] Its further technical solution is: the generalized efficient layer aggregation network structure includes a CSPDenseNet component and an efficient layer aggregation network.

[0014] A further technical solution is that the marking result carries a label of a defect category, and the defect category includes cracks, finger electrode interruptions, black cores and thick lines.

[0015] The present invention also provides a photovoltaic cell defect detection device, comprising: An image acquisition unit, used to acquire an EL image of the photovoltaic cell to be detected; A detection unit, used for inputting the photovoltaic cell EL image to be detected into a photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is a model obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network; An output unit is used to output the detection result.

[0016] The present invention further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the above method when executing the computer program.

[0017] The present invention also provides a storage medium, wherein the storage medium stores a computer program, and the computer program implements the above method when executed by a processor.

[0018] Compared with the prior art, the present invention has the following beneficial effects: the present invention solves the problem of insufficient data set by using StyleGAN3-T to generate an adversarial network to expand the EL image of photovoltaic cells; preprocesses, labels and enhances the expanded image to enhance data diversity and improve the robustness of the model; uses a deep learning network for training so that the model can automatically identify and classify defects in photovoltaic cells and reduce manual intervention and errors; in a complex environment, the model obtained through training can improve the detection accuracy and adaptability and avoid the limitations of traditional methods; the output detection results not only improve the efficiency, but also have better scalability and are suitable for photovoltaic cell monitoring tasks of different types and scales. The whole method solves the problems that the existing photovoltaic monitoring technology mainly relies on traditional image processing methods, which are inefficient, error-prone and difficult to expand, and even machine learning and deep learning technologies still face the challenges of small data sets, poor versatility and insufficient performance in complex and changing environments.

[0019] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.

[0021] Figure 1 A schematic diagram of a process for detecting photovoltaic cell defects provided by an embodiment of the present invention; Figure 2 A schematic diagram of an expanded photovoltaic cell EL image provided by an embodiment of the present invention; Figure 3 A schematic diagram of an average precision index provided by an embodiment of the present invention; Figure 4 A schematic diagram of the precision and recall rate indicators provided by an embodiment of the present invention; Figure 5 A schematic diagram of a confusion matrix provided by an embodiment of the present invention; Figure 6 A schematic diagram of the prediction result of the verification data provided by an embodiment of the present invention, the real data / original label predicted label; Figure 7 A schematic diagram of a prediction result on a test data image provided by an embodiment of the present invention; Figure 8 A schematic diagram of prediction results of unseen data obtained from public sources provided by an embodiment of the present invention; Fig. 9 A schematic diagram of the prediction results of the unseen data of EL imaging in the accelerated aging experiment provided by the embodiment of the present invention; Fig.10 A schematic diagram of an unseen data source and image description for verification provided by an embodiment of the present invention; Fig.11 A schematic diagram showing a comparison between the photovoltaic cell defect detection method provided by an embodiment of the present invention and other frameworks; Fig.12 A schematic block diagram of a photovoltaic cell defect detection device provided by an embodiment of the present invention; Fig.13 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0022] 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 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.

[0023] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0024] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0025] It should be further understood that the term "and / or" used in the present description and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0026] See also Figure 1 , Figure 1 A schematic flow chart of a photovoltaic cell defect detection method provided in an embodiment of the present invention. The photovoltaic cell defect detection method is applied on a server to solve the problems that the existing photovoltaic monitoring technology mainly relies on traditional image processing methods, which are inefficient, prone to errors and difficult to expand, and that even machine learning and deep learning technologies still face challenges such as small data sets, poor versatility and insufficient performance in complex and changing environments.

[0027] Figure 1 FIG. 1 is a flow chart of a photovoltaic cell defect detection method provided by an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S130.

[0028] S110, obtaining an EL image of the photovoltaic cell to be inspected.

[0029] In this embodiment, the EL (electroluminescence) image of the photovoltaic cell to be inspected is usually obtained to evaluate the quality and health of the photovoltaic cell. EL imaging can be used to detect defects such as microcracks and broken grid lines, which are difficult to detect in conventional visual inspections but will affect the working efficiency and life of the photovoltaic cell.

[0030] S120. Input the photovoltaic cell EL image to be detected into the photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is a model obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network.

[0031] In this embodiment, the detection result refers to the defects existing in the photovoltaic cell and the corresponding types.

[0032] A dataset of electroluminescence (EL) images of multicrystalline photovoltaic cells is enhanced by generative artificial intelligence techniques. These steps cover data collection, data generation based on generative AI, preprocessing, labeling, mosaic enhancement, training / optimization, and testing / evaluation. The method of this embodiment targets 3500 EL images of multicrystalline photovoltaic cells containing four defect types: cracks, finger interruptions, black cores, and thick lines.

[0033] In one embodiment, the photovoltaic cell defect detection model is obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network to obtain the model, including steps S121~S126.

[0034] S121, collecting original photovoltaic cell EL images.

[0035] In this embodiment, the original photovoltaic cell EL image can be captured using an infrared camera or a modified digital camera specifically used for capturing EL images, and an appropriate power source to excite the photovoltaic cell to emit light to capture the EL image of the photovoltaic cell.

[0036] S122. Use StyleGAN3-T to generate an adversarial network to expand the original photovoltaic cell EL image to obtain an expanded photovoltaic cell EL image.

[0037] In this embodiment, the expanded photovoltaic cell EL image refers to the result obtained by expanding the original photovoltaic cell EL image using the StyleGAN3-T generative adversarial network.

[0038] In one embodiment, the above-mentioned step S122 may include steps S1221 - S122.

[0039] S1221. Setting a seed to initialize the random generator of the StyleGAN3-T generative adversarial network and setting a set cutoff value; S1222. Based on the original photovoltaic cell EL image, a StyleGAN3-T generative adversarial network is used to generate a synthetic image to obtain an expanded photovoltaic cell EL image; Among them, the StyleGAN3-T generative adversarial network is a pre-trained model that is trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

[0040] In this embodiment, in order to expand the original data set and improve the generalization ability of the model, an advanced image generation AI technology, Style Generative Adversarial Network (StyleGAN), is used. Unlike the traditional generative adversarial network (GAN), StyleGAN introduces an adaptive normalization (AdaIN) mechanism, which transforms the latent vector into a more complex representation through a mapping network, and adds multiple noise inputs to the synthesis network to increase the randomness of image details. In order to explore the best image generation strategy, this paper uses different versions of StyleGAN, such as StyleGAN2, StyleGAN3-T and StyleGAN3-R, among which StyleGAN3-T is selected as the final model due to its advantages in clarity and realism of image details.

[0041] Specifically, the generative adversarial network consists of two neural networks: the generator and the discriminator. The input of the generator network is random (Gaussian) noise, which is used as a latent vector Z to generate images through the generator network P_Z (Z)G, while the discriminator network D is used to distinguish the generated images from the real images. By calculating the back propagation loss of the two networks, the Min-Max loss function is used for optimization. The loss function expression is: .

[0042] In this process, the discriminator needs to maximize its output value so that its judgment is as close to 1 as possible; while the generator needs to minimize its output value so that the discriminator's judgment on the generated image is close to 0.

[0043] In ordinary GAN, random noise is directly fed to the generator as a latent vector Z; in StyleGAN, this process is implemented through a mapping network that uses adaptive normalization (AdaIN) and affine transformation (a). The mapping network converts the latent vector w into a weight in the latent space and inputs it into the AdaIN module together with the convolutional features x of the previous layer. The formula of AdaIN is: It combines batch normalization and instance normalization, and Control, is the weight of the potential vector w. Among them, and The scale and translation of the image are controlled respectively, so AdaIN effectively controls the style of the image. In addition, StyleGAN also introduces multiple noise inputs in the synthesis network to increase the random details of the generated image. StyleGAN2 introduced improved normalization techniques, progressive training strategies, and perceptual path length regularization on the original basis, which effectively solved the artifact problem. StyleGAN3 further optimized StyleGAN2, especially solving the problems of spatial encoding information leakage and texture adhesion. It is called AliasFreeGAN (aliasing-free GAN). Its main innovations include replacing Fourier features to measure input, reducing the width of the mapping network, removing output jump connections, designing higher quality filters, and introducing new learning affine layers.

[0044] In this example, different versions of StyleGAN (StyleGAN2, StyleGAN3-T, and StyleGAN3-R) were trained from scratch, and the learned weights were transferred through a large number of experiments to ensure the high quality of the generated images. The results show that StyleGAN3-T is superior in terms of detail, clarity, and realism of the generated images. To evaluate the effectiveness of these models, the FID (Fréchet Inception Distance) metric was used, which evaluates the realism and fidelity of the generated images by comparing the distribution of the generated images with the distribution of the real images.

[0045] After a lot of experiments, this embodiment selected StyleGAN3-T for training, and trained the 512×512 PV cell EL image dataset through transfer learning. First, the StyleGAN dataset tool was used to convert the images into a format suitable for training. These images were stored as uncompressed PNG files with a file size of a power of 2, which facilitated efficient decoding during training. The transfer learning used the AFHQv2 512×512 pre-trained model, which was originally trained on the animal face dataset. Transfer learning helps to accelerate the training process and improve learning results.

[0046] In the generative model, the seed and truncation values ​​are controlled to adjust the characteristics of the generated images. The seed is used to initialize the random generator, affecting the characteristics and style of the generated images; the truncation controls the balance between the fidelity and diversity of the generated data. In this experiment, the truncation value range is set between 0.5 and 1.5. Some of the generated synthetic images are shown in Figure 1. Figure 2 The generated data is not only used to enhance the original dataset, but will also be used in subsequent model training.

[0047] The method of this embodiment does not rely solely on traditional enhancement methods such as simple image transformation or flipping, but creates new, high-quality EL images through complex generative models, thereby greatly expanding the available data resources and contributing to the development of more accurate and robust photovoltaic cell defect detection systems.

[0048] S123 , preprocessing and marking the expanded photovoltaic cell EL image to obtain a marking result.

[0049] In this embodiment, the marking result is labeled with defect categories, which include cracks, finger electrode interruptions, black cores, and thick lines.

[0050] Specifically, the generated images were used to augment and preprocess the existing dataset. The final images were obtained after removing duplicate or highly similar samples from the dataset, a process designed to improve model performance and avoid potential biases. All images were resized to 640×640. The processed images show polycrystalline cell features of different types, intensities, orientations, compositions, interactions, and brightness, making the dataset very suitable for training models with good robustness. Since polycrystalline cell images usually have more complex background textures, they are more difficult to detect than single crystal cell images.

[0051] These images are annotated into four categories: cracks, interruptions of finger electrodes, black cores, and thick lines. Star-shaped cracks are also classified as cracks. During the annotation process, the labelImg tool was used for annotation, and all annotations were saved in Pascal-VOC or YOLO format as needed. The image content covers various types of cracks (such as linear cracks, multi-directional cracks, and star-shaped cracks), black cores, interruptions of finger electrodes, finger electrodes interrupted along cracks, and thick lines. This dataset faces the problem of class imbalance, so Distribution Focal Loss is used to deal with this problem. Distribution Focal Loss is a variant of Focal Loss that is designed to handle training data with class imbalance, allowing the model to maintain good performance when dealing with rare categories.

[0052] S124: Perform data enhancement on the labeling result to obtain a sample set.

[0053] In this embodiment, the sample set refers to images obtained after data enhancement is performed on the labeled images.

[0054] Specifically, Mosaic is used to perform data enhancement on the labeling result to obtain a sample set.

[0055] In this embodiment, mosaic technology is used for data augmentation and spliced ​​during training by a data loader. Specifically, mosaic combines four different images into a new image to simulate diverse compositions and interactions, thereby providing richer contextual information.

[0056] In addition to the mosaic technique, other common data enhancement methods were used in different experiments, including flipping, scaling, rotation, shearing, translation, contrast adjustment, brightness adjustment, cropping, noise addition and erasing, etc. In addition, some other enhancement methods were explored in the experiments, such as image mixing, copy-paste and shearing techniques.

[0057] S125. Build a deep learning network.

[0058] In this embodiment, the deep learning network includes a Yolov9 model with programmable gradient information (PGI) and a generalized efficient layer aggregation network structure (GELAN).

[0059] The generalized efficient layer aggregation network structure includes a CSPDenseNet component and an efficient layer aggregation network.

[0060] Programmable gradient information and generalized efficient layer aggregation network structure improve learning ability and retain key information. In particular, it solves the common information loss problem in deep neural networks and achieves effective target detection in a lightweight model.

[0061] PGI solves the problem of information loss by introducing the concept of reversible functions, ensuring that gradients can be reliably generated even when the network is deepened, thereby promoting accurate network updates and improved detection performance. Specifically, reversible functions allow information not to be lost during data conversion, providing the model with more reliable gradients and promoting effective training and convergence of the network.

[0062] In this embodiment, in the YOLOv9 version, the problem of poor convergence in deep neural networks due to gradient saturation or vanishing is solved. Although modern deep networks have alleviated this problem to a certain extent through normalization and activation functions, the problem of slow convergence or poor convergence still exists. According to the information bottleneck principle, the loss of important information in the deep network may be the cause of this problem, that is, information is gradually lost during the data conversion process. Specifically, ; where I represents mutual information, f and g represent transformations, and, and is a parameter. The formula shows that as the number of network layers increases, some information of the original data may be lost. This information loss will affect the model's ability to retain the predicted target, which in turn affects the calculation of the loss function and the update of the gradient, ultimately leading to poor convergence performance.

[0063] In order to solve the problem of unreliable gradient and poor convergence, reversible functions can be used. There is an inverse transformation function When , it is said to be a reversible function. ; among them and and They are and Parameters. Using a reversible function, to perform the conversion without losing information, ; This provides more reliable gradient generation for model updates because the transformations consist of reversible functions. This reversibility property is ; Where l represents the lth layer of the Pre-Act residual network (ResNet), and f is the transformation.

[0064] In addition, mask modeling is used. For this purpose, an approximate method is used to find the function The inverse transformation function , so that the transformed features can retain enough information using sparse features, ; Where M is a dynamic binary mask.

[0065] However, when this approach is applied, lightweight models often result in over-compression of parameters, which affects This information bottleneck phenomenon can be To indicate that, Usually only For a small part, the information loss in the feed-forward stage may not be significant. However, it is crucial for lightweight models to lose important information in the feed-forward stage because they have been parameterized.

[0066] To solve this problem, a new training method based on the Programmable Gradient Information (PGI) framework was introduced in YOLOv9. This method includes three key parts: the main branch, the reversible auxiliary branch, and the multi-level auxiliary information. The reasoning process only needs to use the main branch, so there is no additional reasoning overhead. The auxiliary reversible branch helps to generate more reliable gradients and update network parameters.

[0067] In this embodiment, the above-mentioned deep learning network also includes a generalized efficient layer aggregation network structure GELAN, which is an architectural advancement that integrates the cross-stage partial network (CSPNET) block and dense connection design. This structure not only avoids the duplication of gradient information, but also enhances the adaptability of the model to different application ranges. GELAN adopts the CSPDenseNet block as its computing module. The CSPDenseNet block includes a combination of dense blocks and transition layers, in which the feature maps are segmented at the base layer and passed through multiple transition layers. Specifically, at each stage, the output and input of the dense block are connected and integrated through the transition layer to generate the final output. The feature map in the base layer of the dense block is transmitted through the channel Here, the output of the dense layer First, it undergoes a transition. Second, the output of the transition layer and Connect and go through another transition, then generate an output .

[0068] The equation for feedforward transfer in CSPDenseNet is as follows: ; ; ; The weight update equation in CSPDenseNet is as follows: ; ; ; This shows that the gradient from the dense layer and the mapping without dense layer are integrated separately. This avoids duplication of gradient information.

[0069] This process ensures that the gradient information and input feature maps (especially the gradients from the dense layer) are integrated separately, thereby avoiding repeated calculation of gradients and improving the efficiency of the network.

[0070] By combining the CSPNET module with the Efficient Layer Aggregation Network (ELAN), GELAN not only inherits the advantages of both, but also can flexibly combine different computing modules to adapt to a variety of application scenarios. In the target detection task, the YOLOv9 model was trained using the GELAN-e architecture and achieved remarkable results when processing the EL image dataset (YOLO format). During the training process, the SGD optimizer and linear learning rate scheduler were used, and L2 regularization and multiple loss functions (box loss, classification loss, and distribution focus loss) were applied to optimize the network performance.

[0071] Yolov9 GELAN-e was trained on the processed EL image dataset (yolo format) with batch size 6, initial learning rate 0.05 and fractional initial learning rate 0.1. Fractional is the learning rate adjusted relative to the epoch to get the final learning rate. Learning rate warmup was performed in the first three epochs to stabilize the training process. The momentum and learning rate of the bias parameter during the warmup phase were 0.8 and 0.1 respectively. The model utilized a linear learning rate scheduler to adjust the learning rate in a linear pattern relative to the epoch. Cosine learning rate scheduler was also studied but it did not help much in convergence. The model was trained for 200 epochs and the data loader mosaic was set to be turned off during the last 10 epochs. It disabled the mosaic operation in the last 10 epochs. The image size chosen for the input data was 640×640. The SGD optimizer was used as it affects the speed and stability of the model convergence. The momentum factor 0.937 was chosen to affect the gradient update. An L2 regularization weight decay of 0.0005 was used.

[0072] The box loss, classification loss, and distribution focal loss are used to evaluate the predicted values ​​relative to the actual values ​​during network training to guide network optimization. The weights of the box loss component, classification loss, and distribution focal loss are 7.5, 0.5, and 1.5, respectively. The distribution focal loss used here is a variant of the focal loss that is used to deal with unbalanced training data so that the model performs well on rare classes. Focal loss is widely used in object detection tasks to handle unbalanced data by adjusting the loss contribution. It reduces the influence of the main class objects and distinguishes easy samples from difficult samples to improve the cross entropy loss. The data used has the class imbalance problem mentioned earlier, and this loss is used to solve this problem.

[0073] Multiple performance evaluation metrics are used to evaluate the performance of the proposed method. Precision, Recall, and F1-Score are used to measure the ability of the model to detect defects in EL images. These metrics are based on some other basic well-known concepts such as true positive (TP), false positive (FP), false negative (FN), and true negative (TN). IoU is another important metric that quantifies the overlap between the predicted and actual ground truth bounding boxes, such as shown.

[0074] Based on TP, FP, TN and FN, Precision, Recall or True positive rate or sensitivity and F1-score are: ; ; .

[0075] Average Precision (AP) and Mean Average Precision (mAP) are other metrics of great significance in object detector evaluation and provide insights into the trade-off between precision and recall by achieving a harmonious balance between false positives and negatives. The area under the precision-recall curve provides the AP for each defect class separately. mAP is obtained by averaging over all classes, ; can also be calculated as; Where n is the number of classes, is the AP for class k. mAP considers various IoU thresholds. The values ​​considered in the results are obtained when specifying the threshold δ = 0.5iou, which is a commonly used value.

[0076] The method was applied to 3500 EL images of polycrystalline photovoltaic cells with four defects: crack, finger, black core and thick line, and the following results were obtained: The YOLOv9 GELAN-e network with generative AI augmented data achieved a performance of 0.946 map@0.5. The validation losses (box loss, classification loss, and distribution focus loss) approach zero faster than the corresponding training losses, which shows that the model generalizes well to the validation data. This indicates that the model performs well on unseen data and that it is learning important features from the training set. Here, box_loss represents the loss calculated during bounding box prediction, which represents the model learning to accurately locate objects. clas_loss represents the loss calculated when predicting the defect class, which represents the classification ability of the model. The average precision metric (mAP) with respect to the number of epochs is as follows Figure 3 As shown in the figure. In the figure, the mAP@0.5 value is the threshold δiou= 0.5; mAP@0.5:0.95 means that the mAP exceeds the threshold δiou=0.5 ~ 0.95, with a step size of 0.05. The precision and recall rate indicators related to the epoch number are as follows Figure 4 As shown in the figure, the accuracy rate reaches 92% and the recall rate reaches 90%.

[0077] In addition, the precision-recall, precision-confidence, recall-confidence, and f1-confidence curves are analyzed, which represent the strengths and weaknesses of the detector. The precision-recall relationship is closer to the upper right corner of the graph, showing better performance, with an average AP of 94.6% across all categories at an IoU threshold of 0.5. The precision-confidence relationship shows that a perfect precision of 100% is achieved at a confidence of 0.811. The recall-confidence relationship shows that the recall is 99% at the minimum confidence threshold. The f1-confidence relationship shows that the model has an accuracy of 88% at a confidence of 0.446 across all categories.

[0078] The confusion matrix summarizing the classification performance is as follows Figure 5 As shown, there are actual and predicted defect categories on the horizontal and vertical axes respectively. The values ​​given in the diagonal of the matrix are correct predictions, and the other off-diagonal values ​​are incorrect predictions.

[0079] The prediction results of the proposed method on validation data and the ground truth labels are shown in Figure 6 The prediction results of the proposed method for the test data image are shown in Figure 7 shown.

[0080] The versatility of the method was verified using unseen data from EL imaging in accelerated aging experiments under unknown operating conditions and public data from different regions. The data includes half-cut and full-cut single-crystal and multi-cell EL images with different types of defects. In addition, it also includes artificially generated images (based on generative AI). Figure 8 and Fig. 9 Prediction results of the proposed method for unseen data obtained from public sources and EL imaging performed in accelerated aging experiments are shown, respectively. Fig.10 Some of these images are provided along with their respective sources, regions, and descriptions.

[0081] In order to verify the effectiveness of the method of this embodiment, it is compared with several state-of-the-art deep learning-based object detection networks. Specifically, the experiment includes different variants of networks such as Detectron2, EfficientDet, DetectionTransformer (DETR), Yolov7, Yolov8, and Yolov9, especially the Faster R-CNN model. The experiment first uses the default settings of each network framework as a baseline, and then performs hyperparameter tuning on the defect detection task, adjusting the batch size, learning rate, learning rate decay, weight decay, etc. The model is trained for 100 to 200 epochs to find room for performance improvement. The specific training settings are as follows: Detectron2: Use Faster R-CNN+ResNet50-FPN architecture, train for 150 rounds, and use a learning rate of 0.0005.

[0082] EfficientDet: uses compound scaling and Bidirectional Feature Pyramid Network (BiFPN), trained for 150 epochs, with a learning rate of 0.001 and a batch size of 32.

[0083] DETR: Use a ResNet50 backbone network with self-attention mechanism and bipartite matching, trained for 150 epochs with a learning rate of 0.

[0084] YOLO family: including different versions such as Yolov7, Yolov7-tiny, Yolov8s, Yolov8m and Yolov8l, and extensive comparisons were conducted.

[0085] like Fig.11 As shown, the experimental results show that the method of this embodiment is superior to other detection methods in performance. The verification results show that the method of this embodiment shows better accuracy and generalization ability on unseen data from different sources.

[0086] The method of this embodiment is particularly suitable for automatic classification and detection of electroluminescent image defects of polycrystalline silicon photovoltaic cells. Through verification in multiple public data sources and accelerated aging experiments, the method shows strong practical application potential, especially when dealing with unknown environments and unseen data with different characteristics.

[0087] In order to further improve the detection effect, the Yolov9 detector with programmable gradient information architecture (PGI) and general efficient layer aggregation network (GELAN) technology is used (combined with mosaic enhancement in the data loader). Experimental results show that the combination of Yolov9 and these methods, especially when processing unseen verification data, can converge the verification loss to zero more quickly, indicating that it has stronger generalization ability.

[0088] In addition, transfer learning has also been applied to StyleGAN3 translation equivariant training of polycrystalline photovoltaic electroluminescence images to generate images with multi-defect features. Compared with traditional data enhancement methods, data enhancement based on generative adversarial networks (GANs) significantly improves the performance and robustness of the model. The enhanced images generated by AI are both clear and realistic, and can effectively improve the data diversity in the training process, making the proposed method have a wider application prospect.

[0089] In terms of data enhancement, advanced techniques such as mosaic, blending, copy-paste, and cut are used. Experiments show that mosaic enhancement has a significant effect on the model training of electroluminescent images. The data loader performs image mosaics during the training process, synthesizing four different images into one, which can simulate different compositions and interactions, thereby providing richer contextual information and further improving the learning effect of the model.

[0090] S126. Use the sample set to train the deep learning network to obtain a photovoltaic cell defect detection model.

[0091] In this embodiment, an optimized photovoltaic cell defect detection model obtained after training can identify defect areas in electroluminescent images.

[0092] S130: Output the detection result.

[0093] In this embodiment, the trained detection model is applied to the actual photovoltaic cell image to output the defect detection result. The detection result usually includes: Defect location: The model will mark the location of the defect area in the electroluminescence image, usually represented by a bounding box (such as a rectangular box); Defect class or severity: If the model is trained to classify different types of defects (e.g., cracks, hot spots, etc.), the defect type for each inspection area will be output; Confidence score: For each detection result, the model may also output a confidence score, indicating the reliability of the detection result.

[0094] The output results can be presented through image annotation, statistical reports or other methods to facilitate subsequent analysis or further processing.

[0095] The above-mentioned photovoltaic cell defect detection method solves the problem of insufficient data set by using StyleGAN3-T to generate an adversarial network to expand the photovoltaic cell EL image; preprocesses, labels and data enhances the expanded image to enhance data diversity and improve the robustness of the model; uses a deep learning network for training so that the model can automatically identify and classify defects in photovoltaic cells, reducing human intervention and errors; in complex environments, the model obtained through training can improve detection accuracy and adaptability and avoid the limitations of traditional methods; the output detection results not only improve efficiency, but also have better scalability and are suitable for photovoltaic cell monitoring tasks of different types and scales. The whole method solves the problems that the existing photovoltaic monitoring technology mainly relies on traditional image processing methods, which are inefficient, error-prone and difficult to expand, and even machine learning and deep learning technologies still face the challenges of small data sets, poor versatility and insufficient performance in complex and changing environments.

[0096] Fig.12 is a schematic block diagram of a photovoltaic cell defect detection device 300 provided in an embodiment of the present invention. Fig.12 As shown, corresponding to the above photovoltaic cell defect detection method, the present invention also provides a photovoltaic cell defect detection device 300. The photovoltaic cell defect detection device 300 includes a unit for executing the above photovoltaic cell defect detection method, and the device can be configured in a server. Fig.12 The photovoltaic cell defect detection device 300 includes an image acquisition unit 301, a detection unit 302 and an output unit 303.

[0097] The image acquisition unit 301 is used to acquire the EL image of the photovoltaic cell to be detected; the detection unit 302 is used to input the EL image of the photovoltaic cell to be detected into the photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network to obtain the obtained model; the output unit 303 is used to output the detection result.

[0098] In one embodiment, the device further comprises a model training unit, which is used to: Collect original photovoltaic cell EL images; use StyleGAN3-T to generate adversarial networks to expand the original photovoltaic cell EL images to obtain expanded photovoltaic cell EL images; preprocess and label the expanded photovoltaic cell EL images to obtain labeling results; perform data enhancement on the labeling results to obtain a sample set; construct a deep learning network; and use the sample set to train the deep learning network to obtain a photovoltaic cell defect detection model.

[0099] In one embodiment, the model training unit is further used to: A seed is set to initialize a random generator of a StyleGAN3-T generative adversarial network, and a set cutoff value is set; based on the original photovoltaic cell EL image, a synthetic image is generated using the StyleGAN3-T generative adversarial network to obtain an expanded photovoltaic cell EL image; wherein the StyleGAN3-T generative adversarial network is a pre-trained model trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

[0100] In one embodiment, the model training unit is further used to: use Mosaic to perform data enhancement on the labeling result to obtain a sample set.

[0101] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned photovoltaic cell defect detection device 300 and each unit can refer to the corresponding description in the aforementioned method embodiment, and for the convenience and brevity of description, it will not be repeated here.

[0102] The photovoltaic cell defect detection device 300 can be implemented in the form of a computer program. The computer program can be implemented in Fig.13 Runs on the computer device shown.

[0103] See also Fig.13 , Fig.13 5 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 500 may be a server, wherein the server may be an independent server or a server cluster composed of multiple servers.

[0104] See also Fig.13 The computer device 500 includes a processor 502 , a memory and a network interface 505 connected via a system bus 501 , wherein the memory may include a non-volatile storage medium 503 and an internal memory 504 .

[0105] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions, and when the program instructions are executed, the processor 502 can execute a photovoltaic cell defect detection method.

[0106] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500 .

[0107] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a photovoltaic cell defect detection method.

[0108] The network interface 505 is used to communicate with other devices over the network. Fig.13 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 500 to which the solution of the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0109] The processor 502 is used to run the computer program 5032 stored in the memory to implement the following steps: Acquire an EL image of a photovoltaic cell to be detected; input the EL image of the photovoltaic cell to be detected into a photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is obtained by collecting an original photovoltaic cell EL image, and using a StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancing the expanded photovoltaic cell EL image to obtain a sample set, and training a deep learning network to obtain the obtained model; and output the detection result.

[0110] In one embodiment, the processor 502 implements the photovoltaic cell defect detection model by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network to obtain the model step, specifically implementing the following steps: Collect original photovoltaic cell EL images; use StyleGAN3-T to generate adversarial networks to expand the original photovoltaic cell EL images to obtain expanded photovoltaic cell EL images; preprocess and label the expanded photovoltaic cell EL images to obtain labeling results; perform data enhancement on the labeling results to obtain a sample set; construct a deep learning network; and use the sample set to train the deep learning network to obtain a photovoltaic cell defect detection model.

[0111] The deep learning network includes a Yolov9 model with programmable gradient information and a generalized efficient layer aggregation network structure.

[0112] The generalized efficient layer aggregation network structure includes a CSPDenseNet component and an efficient layer aggregation network.

[0113] The labeling results are labeled with defect categories, including cracks, finger electrode interruptions, black cores, and thick lines.

[0114] In one embodiment, when the processor 502 implements the step of using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image to obtain the expanded photovoltaic cell EL image, the processor 502 specifically implements the following steps: Setting a seed to initialize a random generator of the StyleGAN3-T generative adversarial network and setting a set cutoff value; based on the original photovoltaic cell EL image, using the StyleGAN3-T generative adversarial network to generate a synthetic image to obtain an expanded photovoltaic cell EL image; Among them, the StyleGAN3-T generative adversarial network is a pre-trained model that is trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

[0115] In one embodiment, when the processor 502 implements the step of performing data enhancement on the labeling result to obtain a sample set, the processor 502 specifically implements the following steps: Mosaic is used to perform data enhancement on the labeling results to obtain a sample set.

[0116] It should be understood that in the embodiment of the present application, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0117] It can be understood by those skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiment of the above method.

[0118] Therefore, the present invention also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps: Acquire an EL image of a photovoltaic cell to be detected; input the EL image of the photovoltaic cell to be detected into a photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is obtained by collecting an original photovoltaic cell EL image, and using a StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancing the expanded photovoltaic cell EL image to obtain a sample set, and training a deep learning network to obtain the obtained model; and output the detection result.

[0119] In one embodiment, when the processor executes the computer program to implement the photovoltaic cell defect detection model, the steps are as follows: collecting original photovoltaic cell EL images, expanding the original photovoltaic cell EL images using StyleGAN3-T generative adversarial network, preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL images to obtain a sample set, and training the deep learning network to obtain the obtained model steps, specifically implementing the following steps: Collect original photovoltaic cell EL images; use StyleGAN3-T to generate adversarial networks to expand the original photovoltaic cell EL images to obtain expanded photovoltaic cell EL images; preprocess and label the expanded photovoltaic cell EL images to obtain labeling results; perform data enhancement on the labeling results to obtain a sample set; construct a deep learning network; and use the sample set to train the deep learning network to obtain a photovoltaic cell defect detection model.

[0120] The deep learning network includes a Yolov9 model with programmable gradient information and a generalized efficient layer aggregation network structure.

[0121] The generalized efficient layer aggregation network structure includes a CSPDenseNet component and an efficient layer aggregation network.

[0122] The labeling results are labeled with defect categories, including cracks, finger electrode interruptions, black cores, and thick lines.

[0123] In one embodiment, when the processor executes the computer program to implement the step of using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image to obtain the expanded photovoltaic cell EL image, the processor specifically implements the following steps: Setting a seed to initialize a random generator of the StyleGAN3-T generative adversarial network and setting a set cutoff value; based on the original photovoltaic cell EL image, using the StyleGAN3-T generative adversarial network to generate a synthetic image to obtain an expanded photovoltaic cell EL image; Among them, the StyleGAN3-T generative adversarial network is a pre-trained model that is trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

[0124] In one embodiment, when the processor executes the computer program to implement the step of performing data enhancement on the labeling result to obtain a sample set, the processor specifically implements the following steps: Mosaic is used to perform data enhancement on the labeling results to obtain a sample set.

[0125] The storage medium may be a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk, etc., which are computer-readable storage media that can store program codes.

[0126] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0127] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0128] The steps in the method of the embodiment of the present invention can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0129] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.

[0130] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A photovoltaic cell defect detection method, characterized in that: include: Acquire the EL image of the photovoltaic cell to be inspected; The photovoltaic cell EL image to be detected is input into the photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is a model obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network; The detection result is outputted.

2. The photovoltaic cell defect detection method according to claim 1, characterized in that: The photovoltaic cell defect detection model is a model obtained by collecting original photovoltaic cell EL images, expanding the original photovoltaic cell EL images using StyleGAN3-T generative adversarial network, preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL images to obtain a sample set, and training the deep learning network, including: Collect original photovoltaic cell EL images; Using StyleGAN3-T to generate adversarial networks to expand the original photovoltaic cell EL image to obtain an expanded photovoltaic cell EL image; Preprocessing and marking the expanded photovoltaic cell EL image to obtain a marking result; Performing data enhancement on the labeling result to obtain a sample set; Build deep learning networks; The deep learning network is trained using the sample set to obtain a photovoltaic cell defect detection model.

3. The photovoltaic cell defect detection method according to claim 2, characterized in that: The method of using StyleGAN3-T to generate an adversarial network to expand the original photovoltaic cell EL image to obtain an expanded photovoltaic cell EL image includes: Set the seed to initialize the random generator of the StyleGAN3-T generative adversarial network and set the set cutoff value; Based on the original photovoltaic cell EL image, a StyleGAN3-T generative adversarial network is used to generate a synthetic image to obtain an expanded photovoltaic cell EL image; Among them, the StyleGAN3-T generative adversarial network is a pre-trained model that is trained by converting the format of the original photovoltaic cell EL image as a sample set; and during the training process, the Min-Max loss function between the generator and the discriminator is updated, and the similarity between the generated image and the real image is evaluated by calculating the FID score.

4. The photovoltaic cell defect detection method according to claim 2, characterized in that: The step of performing data enhancement on the labeling result to obtain a sample set includes: Mosaic is used to perform data enhancement on the labeling results to obtain a sample set.

5. The photovoltaic cell defect detection method according to claim 2, characterized in that: The deep learning network includes a Yolov9 model with programmable gradient information and a generalized efficient layer aggregation network structure.

6. The photovoltaic cell defect detection method according to claim 5, characterized in that: The generalized efficient layer aggregation network structure includes a CSPDenseNet component and an efficient layer aggregation network.

7. The photovoltaic cell defect detection method according to claim 2, characterized in that: The labeling results are labeled with defect categories, including cracks, finger electrode interruptions, black cores, and thick lines.

8. Photovoltaic cell defect detection device, characterized in that: include: An image acquisition unit, used to acquire an EL image of the photovoltaic cell to be detected; A detection unit, used for inputting the photovoltaic cell EL image to be detected into a photovoltaic cell defect detection model for defect detection and classification to obtain a detection result; wherein the photovoltaic cell defect detection model is a model obtained by collecting the original photovoltaic cell EL image, and using the StyleGAN3-T generative adversarial network to expand the original photovoltaic cell EL image, and then preprocessing, labeling, and data enhancement of the expanded photovoltaic cell EL image to obtain a sample set, and training the deep learning network; An output unit is used to output the detection result.

9. A computer device, characterized in that: The computer device comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images