Photovoltaic cell defect detection method based on S-WIoU loss function

Abstract

The invention provides an S-WIoU loss function-based photovoltaic cell defect detection method, which comprises the following steps of: constructing a defect sample library through a generative adversarial network model, then replacing dynamic weight of a WIoU loss function with a weighting function to form a new WIoU loss function, and replacing IoU loss with the new WIoU loss function to construct an S-WIoU loss function. According to the method, a new WIoU loss function is used as a basis to position loss and adjust a weighting function, so that the sensitivity of a model to different types of small target defects is improved, the small target defects are detected more accurately, then different channel feature responses are adjusted through an SE attention mechanism, and the detection accuracy is improved. The key features of the small target defects are autonomously enhanced, the unnecessary features are automatically inhibited, and the important features of the small target defects are effectively extracted.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology, specifically to a method for detecting defects in photovoltaic cells based on the S-WIoU loss function. Background Technology

[0002] The development and use of clean energy sources such as solar, wind, nuclear, and geothermal energy can alleviate the energy crisis and reduce damage to the ecological environment. In the process of utilizing solar energy, photovoltaic cells are often used to generate electricity. However, during the production and transportation of photovoltaic cells, defects may occur due to uncontrollable objective factors, such as black spots, edge chipping, cross-shaped microcracks, cracks, and perforations. These defects will reduce the power generation efficiency and lifespan of photovoltaic cells. Therefore, researching and improving methods for detecting defects in photovoltaic cells can bring significant value to the advancement of the photovoltaic industry.

[0003] For example, Chinese patent application No. 202410923057.2, published on August 23, 2024, discloses a method for detecting defects in photovoltaic solar cells based on an improved YOLOv5 network. This method utilizes an improved YOLOv5 network, which includes an input terminal, a backbone network, a weighted bidirectional feature pyramid network, and an output terminal. The weighted bidirectional feature pyramid network has three different pathways. These pathways perform multi-scale feature fusion on image features at different levels. The fused features are then connected to a prediction category network and a localization bounding box network, outputting the detected cell defect category and location. Finally, the model is evaluated and the data is analyzed. Applying this model to real-world cell defect detection reduces the impact of the cell's environment on the detection process, accurately identifies the type and location of defects, and detects more small-scale defects with stronger recognition capabilities, thus minimizing missed and false detections.

[0004] The aforementioned literature addresses the issues of insufficient data augmentation and feature fusion in photovoltaic solar cell defect detection by generating data through an improved YOLOv5 network and a BEGAN adversarial network, combined with multi-scale feature fusion using a weighted bidirectional feature pyramid network. However, in the process of detecting small target defects in photovoltaic cells, inaccurate detection results can occur due to deviations in angle, distance, and shape between the detection box and the ground truth box. This leads to the inability to accurately detect small target defects in photovoltaic cells. Furthermore, if the detection box is too large, the number of images of small target defects obtained will be small. Under the premise that the types of small target defects do not continue to increase, it is easy to miss small target defects, thus affecting the convergence speed and training efficiency of the entire detection model. Summary of the Invention

[0005] The purpose of this invention is to provide a photovoltaic cell defect detection method based on the S-WIoU loss function, which can achieve accurate positioning of small target defects, effectively reduce the number of missed small target defects, and improve the robustness of detection.

[0006] To achieve the above objectives, this invention provides a method for detecting defects in photovoltaic cells based on the S-WIoU loss function, comprising the following steps: S1 constructs a defect sample library from the acquired photovoltaic cell defect images using a boundary equilibrium generative adversarial network model to expand the dataset of the defect sample library; S2 constructs the S-WIoU loss function, including steps S2.1 to S2.3. S2.1 The first hyperparameter in the dynamic weights of the WIoU loss function Maximum target area in training set and the true frame area of ​​small target defects Constructing dynamic weight function ; S2.2 will use the weighting function The dynamic weights embedded in the WIoU loss function replace the WIoU loss function to form a new WIoU loss function; S2.3 Combines the new WIoU loss function with the angle loss function, shape loss function, and distance loss function to form the S-WIoU loss function; S3 detects small target defects in the training set by constructing the S-WIoU loss function, adjusts the target defects in the defect sample library, and forms an updated defect sample library; S4 uses a weighted bidirectional feature pyramid network to perform feature fusion on image features at different levels in the updated defect sample library. After feature fusion, the features are verified, and the detected photovoltaic cell defect categories and defect locations are output.

[0007] The above method constructs the S-WIoU loss function. Through the hierarchical collaborative mechanism within the S-WIoU loss function, a new WIoU loss function is formed by first replacing the dynamic weights of the original WIoU loss function with a weight function. This new WIoU loss function then replaces the IoU loss. This allows the new WIoU loss function to serve as the basic localization loss. By adjusting the weight function, the model's sensitivity to small target defects can be effectively improved, resulting in more accurate detection of small target defects. During this process, angle loss, distance loss, and shape loss functions can be added to the S-WIoU loss function, thereby enhancing the detection... The loss function is further adjusted based on the angle difference, distance difference, and shape difference between the bounding box and the ground truth bounding box. This allows the model to improve the accuracy of small target defect detection from the perspective of the detection box. Furthermore, based on the geometric loss composed of angle loss, distance loss, and shape loss, the weight function in the new WIoU loss function is dynamically adjusted to effectively avoid invalid parameters in the search for small target defects, achieve accurate localization of small target defects, and effectively reduce the number of missed detections of small target defects, thereby improving the robustness of detection. At the same time, the boundary model can further improve the initially generated defect sample library, thereby further increasing the database size and making the detection area more extensive.

[0008] Furthermore, the dynamic weight function in step S2.1 As shown in the following formula (1): (1), This is the first hyperparameter in the dynamic weights of the WIoU loss function. This is the second hyperparameter. To maximize the target area in the training set, The true area of ​​the defect in the small target. It is a constant.

[0009] The above settings can construct a new weight function using the first hyperparameter, the second hyperparameter, the maximum target area in the training set, and the ground truth box area of ​​small target defects. By comparing the maximum target area in the training set with the ground truth box area of ​​small target defects, the value of the maximum target area in the training set can be adjusted to obtain different new weight functions, thereby improving the accuracy of detecting different types of target defects.

[0010] Furthermore, the new WIoU loss function in step S2.2 is shown in formula (2) below: (2), For the weight function, This is the traditional crossover loss.

[0011] The above settings can combine the weight function with the WIoU loss function to form a new WIoU loss function.

[0012] Furthermore, the S-WIoU loss function in step S2.3 As shown in the following formula (3): (3), Let be the angle loss function. Let distance be the loss function. This is the shape loss function.

[0013] In the above settings, the S-WIoU loss function replaces the IoU loss with a new WIoU loss function through a hierarchical collaborative machine, and is combined with the constraints of angle loss, distance loss, and shape loss respectively. In this way, the new WIoU loss function can be used as the basic localization loss. By adjusting the weight function, the sensitivity of the model to small target defects can be effectively improved, thereby achieving more accurate detection of small target defects. At the same time, the angle loss prioritizes the correction of the rotation deviation of the bounding box of small target defects, and establishes the optimization path; the distance loss achieves precise fine-tuning of the center point through scale normalization; and the shape loss further constrains the similarity of the aspect ratio of the bounding box of small target defects. Through the combined effect of the geometric loss composed of angle loss, distance loss, and shape loss, invalid parameters for searching small target defects are effectively avoided.

[0014] Furthermore, step S3 also includes: The ground truth area of ​​small target defects in the weighting function during the detection process. When it shrinks, The overall value increases, and then the first hyperparameter is adjusted. Preliminary The overall value is amplified, and then through the second hyperparameter. right The overall value is amplified again to change the hyperparameter values ​​of the S-WIoU loss function.

[0015] The above settings enable small object defects to obtain higher loss weights by changing the area of ​​the ground truth box, thereby allowing the model to dynamically adjust and optimize small object defects during training to achieve accurate detection.

[0016] Furthermore, step S4 also includes step S4.1. When performing feature fusion, S4.1 uses the Squeeze-and-Excitation (SE) attention mechanism to apply global pooling operations to convert the H×W×C dimension spatial feature mapping into a 1×1×C dimension channel descriptor, so as to aggregate the spatial statistical characteristics of different feature channels. The above settings, through channel compression, effectively aggregate the spatial statistical characteristics of each feature channel.

[0017] Furthermore, step S4.1 also includes: during the activation phase, the compressed channel descriptors are processed through two fully connected layers, and then normalized using the Sigmoid activation function to obtain the weights of different channels.

[0018] The above settings allow the compressed channels to be connected and then normalized using an activation function, thus ensuring accuracy after merging multiple channels.

[0019] Furthermore, after step S4.1, there is also step S4.2. In S4.2, the compressed channel descriptors are processed by two fully connected layers through the SE (Squeeze-and-Excitation) attention mechanism, and then normalized by the Sigmoid activation function to obtain the weights of different channels, so as to establish adaptive correlation between different feature channels.

[0020] The above settings utilize the Squeeze-and-Excitation (SE) attention mechanism to dynamically adjust the feature responses of different channels using global context information. This enables the autonomous enhancement of key features of small target defects and the automatic suppression of unnecessary features. By dynamically adjusting the feature regions, the important features of small target defects can be effectively extracted, thereby improving the model's noise resistance.

[0021] Furthermore, The angle loss function is obtained by the following formula (4): (4); The distance loss function is obtained by the following formula (5): (5); The shape loss function is obtained by the following formula (6): (6).

[0022] With the above settings, the angle loss function prioritizes correcting the rotation deviation of the bounding box of small target defects, establishing an optimization path for small target defects; at the same time, the distance loss in the S-WIoU loss function can achieve precise fine-tuning of the center point of small target defects through scale normalization; and the shape loss in the S-WIoU loss function can further constrain the similarity of the aspect ratio of the bounding box of small target defects. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the working method of the present invention.

[0024] Figure 2 This is a schematic diagram of defects in photovoltaic cells in this invention.

[0025] Figure 3 This is a comparison chart showing the detection performance of the +BiFPN+SE+S-WIoU model used in this invention and the original YOLOv5 network model on photovoltaic solar cells.

[0026] Figure 4 This is a schematic diagram illustrating the determination of angle loss in this invention.

[0027] Figure 5 This is a schematic diagram illustrating the determination of distance loss in this invention.

[0028] Figure 6 This is a diagram of the BiFPN network structure in this invention. Detailed Implementation

[0029] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0030] like Figure 1-6 As shown, a method for detecting defects in photovoltaic cells based on the S-WIoU loss function includes the following specific steps: S1 uses a boundary-equilibrium generative adversarial network (BEGAN) model to construct a defect sample library from the acquired photovoltaic cell defect images, thereby expanding the defect sample library.

[0031] S2 constructs the S-WIoU loss function, including steps S2.1 to S2.3. S2.1 The first hyperparameter in the dynamic weights of the WIoU loss function Maximum target area in training set and the true frame area of ​​small target defects Constructing dynamic weight function As shown in the following formula (1): (1), in This is the first hyperparameter in the dynamic weights of the WIoU loss function. This is the second hyperparameter. and The value range of is [0,1]. To maximize the target area in the training set, The true area of ​​the defect in the small target. The first hyperparameter, the second hyperparameter, the maximum target area in the training set, and the ground truth box area of ​​small target defects are used as constants to construct a new weight function. By comparing the maximum target area in the training set with the ground truth box area of ​​small target defects, the value of the maximum target area in the training set can be adjusted to obtain different new weight functions, thereby improving the accuracy of detecting different types of target defects. S2.2 will use the weighting function The dynamic weights embedded in the WIoU loss function replace the WIoU loss function to form a new WIoU loss function, as shown in the following formula (2): (2), in For the new WIoU loss function, The traditional crossover ratio loss is obtained through the traditional IoU loss function, and its specific value can be obtained by calculating 1-IoU. S2.3 Embed the new WIoU loss function into the SIoU loss function to replace the IoU loss component of the SIoU loss function, forming the S-WioU total loss function. As shown in the following formula (3): (3), in Let be the angle loss function. Let distance be the loss function. Using the shape loss function, by constructing the S-WIoU loss function, the hierarchical collaborative mechanism within the S-WIoU loss function can be leveraged. Specifically, the new WIoU loss function replaces the IoU loss, and the angle loss, distance loss, and shape loss optimized for geometric structures are integrated with the dynamically adjusted weight function based on the maximum target area in the training set within the new WIoU loss function. This allows the new WIoU loss function to serve as the basic localization loss. By adjusting the weight function, the model's sensitivity to small target defects can be effectively improved, resulting in more accurate detection of small target defects. This process allows for... The angle loss method prioritizes correcting the rotational deviation of the bounding box of small target defects, establishing an optimized path for small target defects. Meanwhile, the distance loss method can achieve precise fine-tuning of the center point of small target defects through scale normalization. The shape loss method can further constrain the similarity of the aspect ratio of the bounding box of small target defects. Based on the geometric loss composed of angle loss, distance loss and shape loss, the weight function in the new WIoU loss function is dynamically adjusted to effectively avoid searching for invalid parameters of small target defects, achieve accurate localization of small target defects, effectively reduce the number of missed detections of small target defects, and improve the robustness of detection.

[0032] like Figure 4 and Figure 5As shown, The angle loss function is obtained through the following formulas (4)-(6): (4); (5); (6); ——Real frame B GT The height difference between the center point of the prediction box B and the center point of the prediction box B; —The height difference between the center points of the ground truth bounding box and the predicted bounding box; —Width and height of the predicted bounding box and the ground truth bounding box, and These are preset parameter values, where t is the value point, x and y are the x and y coordinates of the predicted bounding box and the ground truth bounding box, and w and h are... Figure 6 The width and length between the predicted bounding box and the ground truth bounding box.

[0033] The ground truth area of ​​small target defects in the weighting function during the detection process. To reduce the size or by using the maximum target area in the training set within the weight function Enlargement enables The overall value increases, and then through the first hyperparameter Preliminary The overall value is amplified, and then through the second hyperparameter. right The overall value is amplified again to change the value of the S-WIoU loss function.

[0034] S3 detects small target defects in the training set by constructing the S-WIoU loss function, adjusts the target defects in the defect sample library and forms an updated defect sample library, which includes the defect location and defect category.

[0035] S4 uses a weighted bidirectional feature pyramid network (BiFPN) to fuse image features at different levels in the updated defect sample library. After feature fusion, the image is used to determine and output the defect location and defect category of the photovoltaic cell through a validation set. In this embodiment, the specific validation process of the validation set is determined by comparing the defect location and defect category in the defect sample library with the image after feature fusion. The specific defect location and defect category are determined by the similarity of the feature images. The implementation process of the validation is existing technology and will not be described in detail here. Step S4 specifically includes steps S4.1 to S4.2. When performing feature fusion, S4.1 uses the Squeeze-and-Excitation (SE) attention mechanism and global pooling to convert the H×W×C spatial feature map into a 1×1×C channel descriptor, thereby aggregating the spatial statistical properties of different feature channels. The compressed channel descriptor is then processed using two fully connected layers and normalized using the Sigmoid activation function to obtain the weights of different channels, thereby establishing adaptive correlation between different feature channels. S4.2 utilizes the Squeeze-and-Excitation (SE) attention mechanism to dynamically adjust the feature responses of different channels using global context information. This enables autonomous enhancement of key features of small target defects and automatic suppression of unnecessary features. After dynamically adjusting the feature regions, it can effectively extract important features of small target defects to improve the model's noise resistance. After feature fusion, the features are connected to the prediction category network and the localization bounding box network, respectively, to output the detected photovoltaic cell defect category and defect location.

[0036] In this embodiment, as Figure 2 As shown, the defect sample library includes six different categories of photovoltaic cell defects: black spots, cracks, subfissures, cruciate recesses, thesis, and chipping. The original dataset contained 798 images of these six types of defects, including 93 cracks, 117 subfissures, 151 chipping images, and 226 cruciate recesses. Among the smaller defects, there were 152 black spots and 59 thesis images. The defect sample library was expanded using the Generative Adversarial Network (BEGAN) model in step S1. This enhances the diversity of defect samples, alleviates data imbalance, and improves overall performance. The expanded photovoltaic cell defect samples are shown in Table 1 below. Table 1

[0037] Images with significant defect features were synthesized, resulting in 282 images of crack defects, 303 images of chipped edges defects, 313 images of black spot defects, 294 images of perforation defects, 320 images of strip-shaped hidden crack defects, and 311 images of cross-shaped hidden crack defects.

[0038] Steps S2 to S4 were performed in the expanded defect sample library. The output data were compared and analyzed through ablation experiments. The data comparison is shown in Table 2 below. Table 2

[0039] Data comparison and analysis combined Figure 3 Compared to the baseline YOLOv5 model, the +BiFPN+SE+S-WioU model (i.e., the model in this embodiment) achieved significant improvements in precision, recall, overall performance metrics, and mAP@0.5% on the dataset with the expanded defect sample library. The complete detection model of +BiFPN+SE+S-WioU, which integrates the weighted bidirectional feature pyramid network (BiFPN), the SE (Squeeze-and-Excitation) attention mechanism, and the S-WIoU loss function, can improve the F1 score to 89.37, increase the recall by 19.6%, and achieve mAP@0.5% of 92.10%. In this embodiment, the F1 value is the harmonic mean of precision and recall, which can convert the model's performance into a quantitative value between 0 and 1.

[0040] To address the issues with the dataset, this paper employs the BEGAN generative adversarial network for data augmentation. To verify the detection performance of the defective images generated by the BEGAN generative adversarial network, this study tested the model's adaptability to different scenarios and data distributions using both augmented and unprocessed initial sample sets, as shown in Table 3, which presents a training comparison.

[0041] Table 3

[0042] Experimental results show that the overall performance of the dataset is significantly improved after data augmentation using BEGAN (the BEGAN model used in this embodiment). Accuracy and recall increased by 0.8% and 7.8%, respectively. With the significant improvement in recall, the BEGAN data augmentation method effectively expanded the feature space of the training samples, promoting the neural network's ability to extract key features. Simultaneously, this method significantly reduced the probability of missed detections, strongly validating the positive role of data augmentation strategies in improving model performance. The detection effect on minor defects is particularly significant, with improvements exceeding 10% in the detection of black spots and hidden cracks. Furthermore, the detection effect of other defects is also greatly improved. The experimental results demonstrate that the data augmentation effect of BEGAN generative adversarial networks is significant, improving the overall detection performance and greatly alleviating the problem of class imbalance in the dataset.

[0043] Among them, the weighted bidirectional feature pyramid network (BiFPN) effectively suppressed background interference through multi-scale feature dynamic weighted fusion, resulting in a slight increase in recall of 0.9%; the Squeeze-and-Excitation (SE) attention mechanism significantly enhanced the representation ability of small target defect features, driving an 8% increase in recall and a 6.4% improvement in mAP@0.5; The above analysis shows that the Squeeze-and-Excitation (SE) attention mechanism enhances the ability to represent small target defects, improving accuracy by 6.4%. The weighted bidirectional feature pyramid network (BiFPN) constructs an adaptive fusion mechanism for defect features, effectively reducing the interference of complex backgrounds on detection results. The S-WIoU loss function dynamically adjusts the weight function, significantly improving the model's sensitivity to small target defects, thereby improving the overall detection accuracy of the model by 6.7%.

[0044] The working principle of this invention is as follows: A defect sample library is constructed using a generative adversarial network (GAN) model. Then, a new WIoU loss function is formed by replacing the dynamic weights of the WIoU loss function with a weight function. This new WIoU loss function is then used to replace the IoU loss to construct an S-WIoU loss function. By using the new WIoU loss function as the base localization loss and adjusting the weight function, the model's sensitivity to different types of small target defects is improved, thus achieving more accurate detection of small target defects. Next, the SE (Squeeze-and-Excitation) attention mechanism is used to adjust the feature responses of different channels, achieving autonomous enhancement of key features of small target defects and automatic suppression of unnecessary features, effectively extracting important features of small target defects. Finally, a weighted bidirectional feature pyramid network is used for feature fusion. The fused features are then compared with a prediction category network. Connected to the localization bounding box network, the system outputs the detected photovoltaic cell defect categories and locations. During this process, the angle loss in the S-WIoU loss function prioritizes correcting the rotational deviation of the small target defect bounding box, establishing an optimized path for the small target defect. Simultaneously, the distance loss in the S-WIoU loss function enables precise fine-tuning of the center point of the small target defect through scale normalization. Furthermore, the shape loss in the S-WIoU loss function further constrains the similarity of the aspect ratio of the small target defect bounding box. Based on the geometric loss composed of angle loss, distance loss, and shape loss, dynamic adjustments are made through the weight function in the new WIoU loss function, effectively avoiding invalid parameters in the search for small target defects, achieving accurate localization of small target defects, effectively reducing the number of missed detections, and improving detection robustness.

Claims

1. A photovoltaic cell defect detection method based on the S-WIoU loss function, characterized in that: Includes the following steps: S1 constructs a defect sample library from the acquired photovoltaic cell defect images using a boundary-equilibrium-based generative adversarial network model to expand the dataset of the defect sample library. S2 constructs the S-WIoU loss function, including steps S2.1 to S2.

3. S2.1 The first hyperparameter in the dynamic weights of the WIoU loss function Maximum target area in training set and the true frame area of ​​small target defects Constructing dynamic weight function ; S2.2 will use the weighting function The dynamic weights embedded in the WIoU loss function replace the WIoU loss function to form a new WIoU loss function; S2.3 Combines the new WIoU loss function with the angle loss function, shape loss function, and distance loss function to form the S-WIoU loss function; S3 detects small target defects in the training set by constructing the S-WIoU loss function, adjusts the target defects in the defect sample library, and forms an updated defect sample library; S4 uses a weighted bidirectional feature pyramid network to perform feature fusion on image features at different levels in the updated defect sample library. After feature fusion, the features are verified, and the detected photovoltaic cell defect categories and defect locations are output.

2. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 1, characterized in that: The dynamic weight function in step S2.1 As shown in the following formula (1): （1）， This is the first hyperparameter in the dynamic weights of the WIoU loss function. This is the second hyperparameter. To maximize the target area in the training set, The true area of ​​the defect in the small target. It is a constant.

3. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 1, characterized in that: The new WIoU loss function in step S2.2 is shown in the following formula (2): （2）， For the weight function, This is the traditional crossover loss.

4. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 1, characterized in that: S-WIoU loss function in step S2.3 As shown in the following formula (3): （3）， Let be the angle loss function. Let distance be the loss function. This is the shape loss function.

5. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 1, characterized in that: Step S3 also includes: The ground truth area of ​​small target defects in the weighting function during the detection process. When it shrinks, The overall value increases, and then the first hyperparameter is adjusted. Preliminary The overall value is amplified, and then through the second hyperparameter. right The overall value is amplified again to change the hyperparameter values ​​of the S-WIoU loss function.

6. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 1, characterized in that: Step S4 also includes step S4.

1. When performing feature fusion, S4.1 uses the Squeeze-and-Excitation (SE) attention mechanism to apply global pooling operations to convert the H×W×C dimension spatial feature mapping into a 1×1×C dimension channel descriptor, so as to aggregate the spatial statistical characteristics of different feature channels.

7. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 6, characterized in that: Step S4.1 also includes: during the activation phase, the compressed channel descriptors are processed through two fully connected layers, and then normalized using the Sigmoid activation function to obtain the weights of different channels.

8. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 6, characterized in that: Step S4.1 is followed by step S4.

2. In S4.2, the compressed channel descriptors are processed by two fully connected layers using the SE (Squeeze-and-Excitation) attention mechanism, and then normalized using the Sigmoid activation function to obtain the weights of different channels, so as to establish adaptive correlation between different feature channels.

9. The photovoltaic cell defect detection method based on the S-WIoU loss function according to claim 4, characterized in that: The angle loss function is obtained by the following formula (4): （4）； The distance loss function is obtained by the following formula (5); （5）； The shape loss function is obtained by the following formula (6): （6）。

Images