Adaptive EL defect detection method based on unsupervised field
Through the unsupervised field adaptive EL defect detection method, virtual data sets are synthesized and the model is trained, which solves the problems of inaccurate segmentation of EL cells and lacks high-quality data sets, and efficient daily monitoring and model migration of photovoltaic modules are achieved.
Patent Information
- Application Number
- CN202510076037.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-08-19
AI Technical Summary
The prior art is difficult to accurately segment EL battery cells, and the lack of high-quality data sets suitable for daily monitoring situations, resulting in low defect detection accuracy and difficult to migrate models to new data sets.
The EL defect detection method that is adaptive to the unsupervised field is adopted. By synthesizing virtual data sets and training unsupervised field adaptive models, Faster R-CNN is used to build student and teacher models, perform adaptive migration of unsupervised fields, and trained in combination with virtual and real data sets.
It effectively reduces the impact of segmentation inaccurate on later identification, solves the problem of lack of high-quality data sets, and realizes daily monitoring of photovoltaic modules and model migration in different environments.
Smart Images
Figure CN120510079A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of defect detection, and in particular to an EL defect detection method based on unsupervised domain adaptation. Background Art
[0002] To achieve the energy transition, the performance and reliability of photovoltaic modules must continue to improve. Therefore, diagnostic tools that assess the extent of module damage during manufacturing, installation, or operation are becoming increasingly important. Electroluminescence (EL) imaging is a fast, non-destructive, and established method for detecting defects in solar modules. While EL images can be acquired very quickly on the ground or from drones, it takes trained professionals 10-30 seconds to analyze each resulting image, which severely limits the number of modules that can be inspected.
[0003] Existing techniques use a method that segments EL images to the cell level before classifying defects. However, different image sets are captured in different environments. This makes it difficult to accurately segment cells when applying the same cell segmentation technology to different image sets, significantly impacting defect detection accuracy. Furthermore, existing defect detection algorithms are mostly designed for EL images of photovoltaic modules captured on production lines or in indoor environments, making these algorithms unsuitable for routine monitoring. Summary of the Invention
[0004] In order to solve the problem that the existing technology is difficult to accurately segment EL cells and rarely considers EL defect detection in daily monitoring situations, an EL defect detection method based on unsupervised domain adaptation is proposed, which can effectively reduce the impact of inaccurate segmentation on subsequent identification and can be widely applied to daily monitoring of photovoltaic modules.
[0005] In order to achieve the above object, the present invention adopts the following technical solution: an EL defect detection method based on unsupervised domain adaptation, comprising the following steps: S1, synthesizing virtual EL component images and annotating EL defects based on the publicly available cell-level EL dataset to obtain a virtual dataset; S2, extract the component area from the captured EL image and perform affine transformation to the front view to obtain the real dataset; S3, trains the unsupervised domain adaptation model based on the virtual dataset and the real dataset, and detects defects in the component-level EL real dataset based on the trained model.
[0006] In this technical solution, a virtual dataset of battery cell synthesis components is adopted. Based on the unsupervised domain adaptation method, the model trained on the virtual dataset is migrated to the real dataset. This can not only effectively reduce the impact of inaccurate segmentation on subsequent recognition, but also effectively solve the problem of lack of high-quality datasets in the field of EL defect detection. At the same time, it also provides a solution for the migration of models between datasets shot in different environments.
[0007] The present invention is further configured such that the process of synthesizing the virtual EL component image includes: Determine the cell arrangement grid diagram of the required component, randomly select defect-free cells and defective cells from the public EL cell classification dataset, and fill the grids of the grid diagram.
[0008] In this technical solution, defective cells and non-defective cells are selected from the public cell-level EL data set and filled into the grid of the network diagram to complete the synthesis process of the virtual EL component image.
[0009] The present invention is further configured as follows: the marking of EL defects is specifically: Generate EL defect annotations based on the location of the defective cell in the grid diagram.
[0010] In this technical solution, after synthesizing the corresponding EL component image, the corresponding annotation is generated.
[0011] The present invention is further configured as follows: Step S2 includes: Extracting component areas based on semantic segmentation, and then performing front view transformation on the component area image; extracting component areas based on semantic segmentation includes: collecting and labeling the component area semantic segmentation dataset, then training and verifying the semantic segmentation model, and finally extracting the component area based on the semantic segmentation model.
[0012] In this technical solution, a semantic segmentation model is constructed to extract component areas, and after the corresponding components are extracted, they are transformed into a front view.
[0013] The present invention is further configured as follows: the collection and labeling of the component area semantic segmentation dataset includes: using a drone to take EL pictures, selecting the picture-labeled component areas to form a training set and a verification set of the semantic segmentation model.
[0014] The present invention is further configured as follows: the training and verification of the semantic segmentation model includes: The images in the training set are input into the semantic segmentation model to obtain the current segmentation result of the model. The loss function is calculated based on the result and manual annotation, and the loss is backpropagated to optimize the model parameters. The model is iterated repeatedly until the model converges, and the model performance is verified on the validation set.
[0015] In this technical solution, the semantic segmentation model is trained and verified to obtain the corresponding model parameters to output the component area.
[0016] The present invention is further configured as follows: extracting component regions based on a semantic segmentation model includes: The other images taken by the drone are input into the converged semantic segmentation model to obtain the segmented component area images.
[0017] In this technical solution, the component area is obtained through the output of the above-mentioned semantic segmentation model.
[0018] The present invention is further configured such that: the unsupervised domain adaptation model specifically uses Faster R-CNN to construct a student model and a teacher model.
[0019] In this technical solution, the weights of the student model are updated through gradient backpropagation, and the weights of the teacher model are the EMA of the student model weights.
[0020] The present invention is further configured as follows: the training process of the unsupervised domain adaptation model includes: Each labeled sample in the virtual dataset is input into the student model after strong data augmentation, and a supervised loss function is obtained based on the output results and labels; Each unlabeled sample in the real dataset is input into the teacher model after weak data augmentation, and then input into the student model after strong data augmentation. The prediction of the teacher model is used as the distillation target of the prediction of the student model. The weights of the student model are updated through gradient backpropagation, and the weights of the teacher model are the EMA of the student model weights.
[0021] In this technical solution, the virtual data set and the real data set are processed accordingly to finally complete the training.
[0022] The present invention is further configured such that the training process of the unsupervised domain adaptation model further comprises: After the training is completed, the teacher model is retained as the model used for the final test.
[0023] In this technical solution, after the training is completed, the student model is discarded and the teacher model is used as the final detection model.
[0024] The present invention can bring the following beneficial effects: The present invention relates to an EL defect detection method based on unsupervised domain adaptation, which can effectively reduce the impact of inaccurate segmentation on subsequent identification and can be widely applied to daily monitoring of photovoltaic modules. The present invention adopts a virtual dataset of battery cell synthesis components, and then migrates the model trained on the virtual dataset to the real dataset based on the unsupervised domain adaptation method, which effectively solves the problem of lack of high-quality datasets in the field of EL defect detection. It also provides a solution for the migration of models between datasets taken in different environments. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 This is a flow chart of an EL defect detection method based on unsupervised domain adaptation in this application.
[0026] Figure 2 This is a schematic diagram of a grid in which a cell is filled into a grid diagram according to an EL defect detection method based on unsupervised domain adaptation in the present application.
[0027] Figure 3 This is a schematic diagram of generating EL defect annotations based on an unsupervised domain adaptive EL defect detection method of the present application.
[0028] Figure 4 This is an example diagram of the annotation of component area extraction in an EL defect detection method based on unsupervised domain adaptation in this application.
[0029] Figure 5 This is a schematic diagram of an unsupervised domain adaptation model of an EL defect detection method based on unsupervised domain adaptation in the present application. DETAILED DESCRIPTION
[0030] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0031] Explanation of EL images, weak data augmentation, and strong data augmentation.
[0032] EL imaging: EL stands for electroluminescence, also known as field-induced luminescence (EL), and can detect defects in module cells for quality control. The basic principle of EL testing is that a forward bias voltage is applied to a crystalline silicon solar cell, injecting a large number of unbalanced carriers into the cell. Electroluminescence relies on the continuous recombination and emission of photons from these unbalanced carriers injected from the diffusion region. These photons are then captured by a CCD camera and processed by a computer to form an image. The entire testing process must be completed in a darkroom, and the emitted light is infrared.
[0033] Weak data augmentation and strong data augmentation: Strong data augmentation refers to a large degree of transformation of the original data, which usually changes some basic characteristics of the data. Weak data augmentation refers to a small degree of transformation of the original data, which usually does not change the basic characteristics of the data.
[0034] The existing technology uses a method of first segmenting the EL image to the cell level and then classifying the defects. However, the shooting environments of different image sets are different, resulting in the same cell segmentation technology being applied to different image sets. It will be difficult to accurately cut the cell, which will have a great impact on the defect detection accuracy.
[0035] In the existing technology, most defect detection algorithms are only used for EL images of photovoltaic modules taken on production lines or in indoor environments. These defect detection algorithms are not suitable for daily monitoring activities.
[0036] At the same time, existing defect detection technologies are generally supervised methods, which have high requirements on the quality and data of the dataset; however, in actual production scenarios, EL image acquisition is difficult, and it is difficult to generate a dataset that meets the requirements; data labeling also consumes a lot of manpower and material resources.
[0037] In addition, existing methods only train models for specific datasets. When the shooting environment of the new dataset is different from that of the original dataset, it is likely that its data distribution will be different from that of the original dataset. At this time, the model will be difficult to migrate to the new dataset, resulting in poor detection results.
[0038] Example 1 In order to solve the problem that the existing technology is difficult to accurately segment EL cells, rarely considers EL defect detection in daily monitoring situations, lacks high-quality data sets, and is difficult to migrate models to new data sets. This embodiment proposes an EL defect detection method based on unsupervised domain adaptation, referring to Figure 1 , which mainly includes the following steps.
[0039] Step S1: synthesize EL component images using a public cell-level EL dataset and generate corresponding annotations to obtain a virtual dataset. This step mainly includes virtual EL component image synthesis and EL defect annotation processes.
[0040] The steps of synthesizing the virtual EL component image are more detailed and mainly include the following processes.
[0041] First, we define the cell layout grid diagram of the required components. Then, we randomly select non-defective cells and defective cells from the public EL cell classification dataset and fill them into the grid of the grid diagram. The specific effect after filling can be referred to Figure 2 .
[0042] In the above technical solution, defective cells and non-defective cells are selected from the public cell-level EL data set and filled into the grid of the network diagram to complete the synthesis process of the virtual EL component image.
[0043] The steps for EL defect marking are more detailed and mainly include the following processes.
[0044] The corresponding EL defect annotation is generated according to the position of the defective cell in the grid diagram obtained during the virtual EL component image synthesis process.
[0045] In the above technical solution, after synthesizing the corresponding EL component image, the corresponding annotation is generated.
[0046] Step S2: capture and obtain an EL image, specifically by capturing it with a drone. Extract the component area from the EL image, then perform an affine transformation to transform it to a front view perspective, and finally obtain a real data set.
[0047] The above-mentioned step S2 mainly includes the following processes.
[0048] First, component regions are extracted based on semantic segmentation, and then front view transformation is performed based on the component region images.
[0049] The above process of extracting component areas is described in more detail and includes the following three sub-steps.
[0050] Sub-step 1: Collect and annotate a component region semantic dataset. Specifically, drones are used to capture EL images. After the images are captured, the annotated component regions are selected to form the training and validation sets for the semantic segmentation model.
[0051] Sub-step 2: Train and validate the semantic segmentation model. More specifically, the images in the training set obtained in Sub-step 1 above are fed into the semantic segmentation model to obtain the model's current segmentation results. Based on these results and the manual annotations, a loss function is calculated, and the loss is back-propagated to optimize the model parameters. Iterations are repeated until the model converges, and the model performance is verified on the validation set.
[0052] In the above technical solution, the semantic segmentation model is trained and verified to obtain corresponding model parameters to output the component area.
[0053] Sub-step 3: Obtain the component region based on the trained semantic segmentation model. More specifically, other images taken by the drone are input into the converged semantic segmentation model to obtain the segmented component region images.
[0054] In the above technical solution, a semantic segmentation model is constructed to extract component areas, and after the corresponding components are extracted, they are subjected to front view transformation.
[0055] In step S3, the unsupervised domain adaptation model is trained by combining the virtual dataset and the real dataset. After the training is completed, the trained model is used to detect defects in the component-level EL real dataset.
[0056] In more detail, the above unsupervised domain adaptation model specifically uses Faster R-CNN to construct the student model and teacher model.
[0057] The training steps of the unsupervised domain adaptation model mainly include the following.
[0058] Each labeled sample in the virtual dataset is input into the student model after strong data augmentation, and a supervised loss function is obtained based on the output results and labels.
[0059] Each unlabeled example in the real dataset is fed into the teacher model after weak data augmentation and then into the student model after strong data augmentation. The teacher model's predictions serve as the distillation target for the student model's predictions. The student model's weights are updated via gradient backpropagation, and the teacher model's weights are the EMA of the student model's weights.
[0060] In the above technical solution, the virtual data set and the real data set are processed accordingly, and finally the training is completed.
[0061] After the training is completed, the student model is discarded and the teacher model is retained as the model used for final detection.
[0062] In this technical solution, a virtual dataset of battery cell synthesis components is adopted. Based on the unsupervised domain adaptation method, the model trained on the virtual dataset is migrated to the real dataset. This can not only effectively reduce the impact of inaccurate segmentation on subsequent recognition, but also effectively solve the problem of lack of high-quality datasets in the field of EL defect detection. At the same time, it also provides a solution for the migration of models between datasets shot in different environments.
[0063] In this embodiment, the process of first segmenting the component area and then directly locating the defect can effectively reduce the impact of inaccurate segmentation in the early steps on the later identification.
[0064] In this embodiment, a drone is used to capture EL images of photovoltaic panels and a corresponding algorithm is set, which can be widely applied to daily monitoring of photovoltaic modules.
[0065] Example 2 This embodiment proposes an EL defect detection method based on unsupervised domain adaptation, referring to Figure 1 , which mainly includes the following steps.
[0066] Step S1: synthesize EL component images using a public cell-level EL dataset and generate corresponding annotations to obtain a virtual dataset. This step mainly includes virtual EL component image synthesis and EL defect annotation processes.
[0067] The steps of synthesizing the virtual EL component image are more detailed and mainly include the following processes.
[0068] First, we define the cell layout grid diagram of the required components. Then, we randomly select non-defective cells and defective cells from the public EL cell classification dataset and fill them into the grid of the grid diagram. The specific effect after filling can be referred to Figure 2 .
[0069] In the above technical solution, defective cells and non-defective cells are selected from the public cell-level EL data set and filled into the grid of the network diagram to complete the synthesis process of the virtual EL component image.
[0070] The steps for EL defect marking are more detailed and mainly include the following processes.
[0071] The corresponding EL defect annotation is generated according to the position of the defective cell in the grid diagram obtained during the virtual EL component image synthesis process.
[0072] In the above technical solution, after synthesizing the corresponding EL component image, the corresponding annotation is generated.
[0073] Step S2: capture and obtain an EL image, specifically by capturing it with a drone. Extract the component area from the EL image, then perform an affine transformation to transform it to a front view perspective, and finally obtain a real data set.
[0074] The above-mentioned step S2 mainly includes the following processes.
[0075] First, component regions are extracted based on semantic segmentation, and then front view transformation is performed based on the component region images.
[0076] The above process of extracting component areas is described in more detail and includes the following three sub-steps.
[0077] Sub-step 1: Collect and annotate a component region semantic dataset. Specifically, drones are used to capture EL images. After the images are captured, the annotated component regions are selected to form the training and validation sets for the semantic segmentation model.
[0078] Sub-step 2: Train and validate the semantic segmentation model. More specifically, the images in the training set obtained in Sub-step 1 above are fed into the semantic segmentation model to obtain the model's current segmentation results. Based on these results and the manual annotations, a loss function is calculated, and the loss is back-propagated to optimize the model parameters. Iterations are repeated until the model converges, and the model performance is verified on the validation set.
[0079] In the above technical solution, the semantic segmentation model is trained and verified to obtain corresponding model parameters to output the component area.
[0080] Sub-step 3: Obtain the component region based on the trained semantic segmentation model. More specifically, other images taken by the drone are input into the converged semantic segmentation model to obtain the segmented component region images.
[0081] In the above technical solution, a semantic segmentation model is constructed to extract component areas, and after the corresponding components are extracted, they are subjected to front view transformation.
[0082] In step S3, the unsupervised domain adaptation model is trained by combining the virtual dataset and the real dataset. After the training is completed, the trained model is used to detect defects in the component-level EL real dataset.
[0083] In more detail, the above unsupervised domain adaptation model specifically uses Faster R-CNN to construct the student model and teacher model.
[0084] The training steps of the unsupervised domain adaptation model mainly include the following.
[0085] Each labeled sample in the virtual dataset is input into the student model after strong data augmentation, and a supervised loss function is obtained based on the output results and labels.
[0086] Each unlabeled sample in the real data set is input into the teacher model after weak data enhancement, and then input into the student model after strong data enhancement. The prediction of the teacher model serves as the distillation target of the student model's prediction.
[0087] In the above technical solution, the virtual data set and the real data set are processed accordingly, and finally the training is completed.
[0088] After the training is completed, the student model is discarded and the teacher model is retained as the model used for final detection.
[0089] Based on the above content, this embodiment proposes a more specific implementation method, referring to Figure 1 , which mainly includes the following steps.
[0090] First, a component-level EL virtual dataset is generated. Specifically, the public cell-level EL dataset is used to synthesize the EL component image and generate corresponding annotations.
[0091] This step mainly includes the virtual EL component image synthesis process and the EL defect annotation generation process.
[0092] For the virtual EL component image synthesis process, the cell arrangement of the component required in this embodiment is 6×10, and each cell size is 40×40 pixels, so a 6-row 10-column grid is generated; defect-free cells and defective cells are randomly selected from the public cell-level EL dataset, scaled to 40×40, and filled into the grid of the grid. The effect after filling can be seen in the following figure. Figure 2 .
[0093] Regarding the EL defect annotation generation process, in this embodiment, the EL defect annotation can be finally generated according to the position of the defective cell in the grid map; if the defective cell is in the kth row and jth column, the corresponding annotation is the four corner points of the cell in the image, namely ((k-1)*40, (j-1)*40), ((k-1)*40, j*40), (k*40, (j-1)*40), (k*40, j*40), for details, please refer to Figure 3 .
[0094] Subsequently, a component-level EL real dataset is generated. Specifically, the component area is segmented using the EL image obtained by drone photography, and then an affine transformation is performed on it to obtain the front view perspective.
[0095] The process mainly includes two processes: component area extraction based on semantic segmentation and component area image front view transformation.
[0096] For component region extraction based on semantic segmentation, in this example, 20 images are selected to annotate component regions, and the training set and validation set of the semantic segmentation model are randomly separated in a 1:1 ratio. For specific annotation examples, please refer to Figure 4 .
[0097] In this embodiment, a U-Net semantic segmentation model is specifically selected. Images from the training set are fed into the model to obtain the model's current segmentation results. A loss function is then calculated based on this result and the manual annotations. The loss is then back-propagated to optimize model parameters. This process is repeated until the model converges, and model performance is verified on a validation set. Additional images captured by the drone are fed into the converged semantic segmentation model to obtain segmented component region images. Components at the edges are filtered out, retaining only intact components.
[0098] Regarding the front view transformation of the component area image, it is specifically to transform the component area image to the viewing angle of the front view through an affine transformation.
[0099] Finally, the unsupervised domain adaptation model is trained and EL defect detection is performed.
[0100] In this embodiment, an unsupervised domain adaptation method is used to train the defect detection model. The overall framework can be referred to Figure 5 The training set consists of a virtual dataset (with labels) and a real dataset (without labels). At each step in the training, a batch of size B contains data from both the virtual dataset and the real dataset, i.e., B = B src +B tgt , where B src represents data from a virtual dataset, B tgt Represents data from real datasets.
[0101] The student-teacher framework is adopted, and Faster R-CNN is selected to build the student model and the teacher model. The initial weights are set to the pre-trained weights on ImageNet. The weight of the student model θ stu By back-propagating the gradient, the weights θ of the teacher model are updated tch is the EMA (Exponential Moving Average) of the student weights, i.e. θ tch =αθ tch +(1-α)θ stu， α∈[0,1].
[0102] During training, for each labeled sample x in the virtual dataset src，i , first through strong data enhancement t~T src , then input the student model, according to the output result and the corresponding label y src，i Computing supervised loss functions The supervised loss function is specifically expressed as: in, This is the target detection loss function of Faster R-CNN.
[0103] During training, for each unlabeled sample x in the real dataset tgt，i , respectively after weak data enhancement and strong data augmentation t~T tgt , and then input the teacher model and student model respectively. The teacher model’s prediction The prediction p of the student model will be tgt，i The distillation target, distillation loss Specifically expressed as: Among them, the output of the teacher model is post-processed as a soft target, i.e. using softmax output, You can choose a commonly used target detection loss function.
[0104] After the training is completed, the student model is discarded and the teacher model is retained as the final detection model.
[0105] Finally, defects in component-level EL real datasets are detected based on the teacher model.
[0106] Example 3 Based on Example 1, the step of extracting component regions based on semantic segmentation can use rotation target detection or traditional digital image processing to extract component regions.
[0107] Example 4 In Example 1, regarding the training of the unsupervised domain adaptation model, specifically, the detection model is migrated from the virtual dataset to the real dataset to realize EL defect detection in the current scenario; of course, this embodiment can also migrate from datasets shot in other scenarios to the current scenario.
[0108] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.
Claims
1. A method for detecting EL defects based on unsupervised domain adaptation, characterized in that: The following steps are involved: S1, synthesizing virtual EL component images and annotating EL defects based on the publicly available cell-level EL dataset to obtain a virtual dataset; S2, extract the component area from the captured EL image and perform affine transformation to the front view to obtain the real dataset; S3, trains the unsupervised domain adaptation model based on the virtual dataset and the real dataset, and detects defects in the component-level EL real dataset based on the trained model.
2. The EL defect detection method based on unsupervised domain adaptation according to claim 1, characterized in that: The process of synthesizing the virtual EL component image includes: Determine the cell arrangement grid diagram of the required component, randomly select defect-free cells and defective cells from the public EL cell classification dataset, and fill the grids of the grid diagram.
3. The EL defect detection method based on unsupervised domain adaptation according to claim 1 or 2, characterized in that: The marked EL defects are specifically: Generate EL defect annotations based on the location of the defective cell in the grid diagram.
4. The EL defect detection method based on unsupervised domain adaptation according to claim 3, characterized in that: The step S2 comprises: Extracting component areas based on semantic segmentation, and then performing front view transformation on the component area image; extracting component areas based on semantic segmentation includes: collecting and labeling the component area semantic segmentation dataset, then training and verifying the semantic segmentation model, and finally extracting the component area based on the semantic segmentation model.
5. The EL defect detection method based on unsupervised domain adaptation according to claim 4, characterized in that: The collection and labeling of the component area semantic segmentation dataset includes: EL images are taken by drones, and the image annotation component areas are selected to form the training set and validation set of the semantic segmentation model.
6. The EL defect detection method based on unsupervised domain adaptation according to claim 4, characterized in that: The training and verification of the semantic segmentation model includes: The images in the training set are input into the semantic segmentation model to obtain the current segmentation result of the model. The loss function is calculated based on the result and manual annotation, and the loss is backpropagated to optimize the model parameters. The model is iterated repeatedly until the model converges, and the model performance is verified on the validation set.
7. The EL defect detection method based on unsupervised domain adaptation according to claim 4, 5 or 6, characterized in that: The component area extraction based on the semantic segmentation model includes: The other images taken by the drone are input into the converged semantic segmentation model to obtain the segmented component area images.
8. The EL defect detection method based on unsupervised domain adaptation according to claim 1, characterized in that: The unsupervised domain adaptation model specifically uses Faster R-CNN to construct the student model and the teacher model.
9. The EL defect detection method based on unsupervised domain adaptation according to claim 1 or 8, characterized in that: The training process of the unsupervised domain adaptation model includes: Each labeled sample in the virtual dataset is input into the student model after strong data augmentation, and a supervised loss function is obtained based on the output results and labels; Each unlabeled sample in the real dataset is input into the teacher model after weak data augmentation, and then input into the student model after strong data augmentation. The prediction of the teacher model is used as the distillation target of the prediction of the student model. The weights of the student model are updated through gradient backpropagation, and the weights of the teacher model are the EMA of the student model weights.
10. The EL defect detection method based on unsupervised domain adaptation according to claim 9, characterized in that: The training process of the unsupervised domain adaptation model also includes: After the training is completed, the teacher model is retained as the model used for the final test.