Dual-network ensemble learning method and system for photovoltaic cell defect detection

Abstract

The present application relates to the field of artificial intelligence and new energy detection technology, in particular to a dual-network ensemble learning method and system for photovoltaic cell defect detection, the present application constructs two CNN models with the same structure, respectively denoted as model A and model B; the cross-entropy loss preheating training is carried out on the dual models; in the collaborative training stage, model A continues to use the cross-entropy loss optimization, and model B uses the Mixup loss optimization based on image interpolation; every fixed period T, the network parameters of the dual models are aggregated and redistributed; after the training is completed, the average dual model is used to output the defect classification of the prediction result of the EL image. Through the heterogeneous loss design and the periodic parameter aggregation mechanism, the present application effectively improves the robustness of the model to the class imbalance and label noise in the training data, and significantly improves the accuracy and reliability of the photovoltaic cell defect detection.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and new energy detection technology, and in particular to a dual-network integrated learning method and system for photovoltaic cell defect detection. Background Technology

[0002] With the increasing global demand for clean energy, the photovoltaic industry has developed rapidly. However, photovoltaic cells are prone to defects such as microcracks and grid breaks during their production, transportation, installation, and long-term operation, severely affecting power generation efficiency and system stability. Therefore, efficient and reliable defect detection is crucial. Traditional photovoltaic cell defect detection methods, such as manual visual inspection and infrared thermal imaging, suffer from low efficiency, high subjectivity, and difficulty in identifying minute defects. In recent years, automatic defect classification technology based on electroluminescence (EL) images using convolutional neural networks (CNNs) has become a research hotspot. EL imaging can clearly present internal defects as dark areas, providing a high-quality data source for automated analysis. However, existing CNN-based defect classification methods still have significant shortcomings.

[0003] First, most methods rely on complex network structures meticulously designed for specific tasks or extensive hyperparameter tuning, resulting in poor flexibility and scalability, making them difficult to integrate directly into existing, mature CNN models. For example, Zhang et al. (“A lightweight network for photovoltaic cell defect detection inelectroluminescence images based on neural architecture search and knowledge distillation”) dynamically generate lightweight CNNs based on techniques such as neural architecture search and knowledge distillation, but this requires extensive hyperparameter tuning and validation set testing, which are not feasible in real-world scenarios. HM AI-Otum et al. (“Defect classification of photovoltaic solar cells using electroluminescence imaging and enhanced deep learning approaches”) improve defect classification performance by designing a special network structure paired with smooth label loss, but this structure is relatively complex and not suitable for binary classification scenarios. Some methods are sensitive to data defects. In practical applications, training data commonly suffers from class imbalance and semantic ambiguity between some samples and their labels. This is mainly because, in binary classification tasks, existing methods often use a fixed threshold (e.g., 0.5) to convert the probability labels obtained from expert voting (e.g., 0.33, 0.67) into hard labels (normal / defective). This can incorrectly label some samples with minor defects as "normal," leading to serious ambiguity between features and labels. For example, Chen Haiyong et al.'s method ("A defect classification method for electroluminescence images of photovoltaic solar cells") achieved some performance improvement by combining feature descriptors with a CNN classifier, but it ignored problems such as sample class imbalance and label ambiguity, making it unsuitable for binary classification scenarios and requiring further performance improvement in multi-class scenarios. H. Tella et al. proposed a multi-network ensemble training method based on a voting mechanism, but the voting mechanism did not fully utilize the advantages and prediction information of multiple models, resulting in performance lagging behind ordinary training methods. Since most existing methods ignore the aforementioned data-level challenges, the trained models are not robust enough, their performance is unstable in real-world scenarios, and they are prone to false positives or false negatives, increasing unnecessary maintenance costs. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a dual-network integrated learning method and system for photovoltaic cell defect detection, so as to solve the problems of insufficient flexibility and scalability caused by the complexity of network structure and strong hyperparameter dependence of existing methods, as well as the poor robustness and weak generalization ability caused by ignoring the class imbalance and label noise (ambiguity) problems that are common in training data.

[0005] To achieve the above objectives, this invention provides a dual-network integrated learning method for photovoltaic cell defect detection, comprising the following steps:

[0006] Step S1: Construct two convolutional neural network models with identical structures, namely Model A and Model B;

[0007] Step S2: Perform pre-training on the training dataset using cross-entropy loss for models A and B;

[0008] Step S3: Perform iterative heterogeneous training on Model A and Model B after preheating training. Each training round includes:

[0009] Step S31: Continue to optimize model A using cross-entropy loss, and optimize model B using Mixup loss based on sample interpolation.

[0010] Step S32: Every fixed T training rounds, aggregate the network parameters of model A and model B, and redistribute the aggregated parameters to the two models.

[0011] Step S4: Repeat steps S31 and S32 for iterative heterogeneous training until the preset total number of training rounds is reached;

[0012] Step S5: Input the electroluminescent image of the photovoltaic cell to be detected into the trained model A and model B, and average the output prediction results of the two models to obtain the defect classification result of the image to be detected.

[0013] Preferably, in step S2, the preheating training lasts for 5 rounds.

[0014] Preferably, in each round of preheating training, the formula for the cross-entropy loss function is:

[0015] ;

[0016] in This represents the cross-entropy loss, where n represents the total number of samples in a training batch. This represents the image of the i-th sample. Let f represent the hard label of the i-th sample, and let f represent the convolutional neural network model. This represents the network parameters of the model.

[0017] Preferably, for each training sample in the training dataset, the Mixup loss is calculated in the following manner:

[0018] Another sample is randomly selected from the training dataset, and a mixing coefficient is sampled from the Beta distribution. The sampled mixing coefficient is used to perform linear interpolation between the image of the current sample and the image of the random sample to generate an interpolated image.

[0019] The interpolated image is input into the model B, and the cross-entropy loss between its output and the label of the current sample and the label of the random sample is calculated.

[0020] The Mixup loss for this training sample is obtained by weighting and summing the two cross-entropy losses using the mixing coefficients.

[0021] Preferably, the hyperparameter of the Beta distribution is set to 1.

[0022] Preferably, in step S32, the method for aggregating the network parameters of model A and model B is as follows:

[0023] ;

[0024] in, and These represent the network parameters of model A and model B, respectively. These are the aggregated global parameters.

[0025] Preferably, in step S5, averaging the output prediction results of the two models includes:

[0026] Calculate the arithmetic mean of the output prediction results of Model A and Model B for the photovoltaic cell electroluminescence image to be detected, and select the category with the highest probability value in the average value as the final classification result of the image to be detected.

[0027] This invention also provides a dual-network integrated learning system for photovoltaic cell defect detection, comprising:

[0028] The model building module is used to build and initialize two convolutional neural network models with the same structure, namely model A and model B;

[0029] The preheating training module is used to preheat training the models A and B using the cross-entropy loss function;

[0030] The collaborative training control module, used to control the iterative training process, includes:

[0031] A heterogeneous training unit is used to optimize model A using cross-entropy loss, while simultaneously co-training model B using Mixup loss.

[0032] The parameter aggregation unit is used to periodically aggregate and redistribute the network parameters of model A and model B according to a preset aggregation period T.

[0033] An iterative control module is used to control the collaborative training control module to run repeatedly until the preset total number of training rounds is reached;

[0034] The defect detection module is used to receive the image to be detected, call the trained model A and model B to make predictions, and average the prediction results to output the defect classification result.

[0035] The beneficial effects of this invention are:

[0036] 1. This invention creatively combines strong fitting ability with noise-robust learning by constructing heterogeneous loss training. Model A focuses on efficiently capturing the overall data distribution, while Model B, through Mixup interpolation, forces the network to learn more stable and semantically invariant features between samples, thereby jointly enhancing the model's inherent resistance to class imbalance and label noise in the training data. This makes the model more stable and reliable in the face of complex and defective data in real industrial scenarios, significantly reducing false positives and false negatives.

[0037] 2. This invention introduces a periodic parameter aggregation and re-initialization mechanism to periodically fuse the parameters of two independently optimized models and reset them to a new starting point for continued collaborative training. This achieves knowledge complementarity and transfer between the two models at the parameter level, effectively integrating the accurate discrimination ability of model A with the robust semantic representation ability of model B, thereby improving the overall performance of the models.

[0038] 3. In the defect detection stage, this invention adopts a strategy of averaging the prediction results of the dual-model soft labels, which can smooth out the abnormal or uncertain predictions that may be generated by a single model, and make a more robust final classification decision by using integrated information, thereby further improving the reliability and confidence of the system in actual deployment.

[0039] 4. This invention, through a mixup and collaborative training mechanism, enables the dual models to fully utilize all available training samples for learning. This avoids ignoring marginal cases or rare defect patterns, ensuring the models can learn from all data, and significantly improves detection capabilities for photovoltaic EL image datasets with small samples or sparse defect types. Attached Figure Description

[0040] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a schematic diagram of the dual-network integrated learning method for photovoltaic cell defect detection according to an embodiment of the present invention;

[0042] Figure 2 This is a block diagram of a dual-network integrated learning system for photovoltaic cell defect detection according to an embodiment of the present invention;

[0043] Figure 3 This is a comparison chart of the defect detection results of the DHA network in this embodiment of the invention with other methods on defect images of monocrystalline silicon and polycrystalline silicon. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0045] It should be noted that, unless otherwise defined, the technical or scientific terms used in this invention should have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0046] Example 1:

[0047] like Figure 1 As shown, this embodiment provides a dual-network ensemble learning method for photovoltaic cell defect detection, assuming a photovoltaic cell electroluminescence (EL) image training dataset. It contains k categories, where n represents the total number of training samples. This represents the i-th input image. This represents its corresponding hard label. Here, H, W, and C represent the height, width, and number of channels of the input image, respectively. The method includes the following steps:

[0048] Step S1: Construct two identical convolutional neural network (CNN) models, model A and model B, forming an overall network framework called DHA. The overall framework is as follows: Figure 2 As shown, for each model Its network parameters are denoted as For a given input sample The output is represented as ;

[0049] Step S2: Perform warm-up training on Model A and Model B using cross-entropy (CE) loss on the training dataset for 5 epochs. In each epoch, the formula for calculating the cross-entropy loss is:

[0050] (1);

[0051] in, This represents the cross-entropy loss, where n represents the total number of samples in a training batch. This represents the image of the i-th sample. Let f represent the hard label of the i-th sample, and let f represent the convolutional neural network model. This represents the network parameters of the model.

[0052] This operation aims to give the two models preliminary discriminative capabilities while ensuring consistency with the subsequent heterogeneous loss-based co-training process. Warming up with CE (Corrective Execution) is beneficial for the subsequent heterogeneous loss-driven training process, potentially leading to performance improvements.

[0053] Step S3: Perform iterative heterogeneous training on model A and model B after preheating training. In the t-th epoch of the iterative heterogeneous training (t starts from 0):

[0054] Step S31: Continue to optimize model A using cross-entropy loss, while replacing the loss function of model B with Mixup loss. When calculating Mixup loss, for each sample pair in dataset D... First, based on the Beta distribution Interpolation is performed on the input image as defined by the following formula:

[0055] (2)

[0056] in, Another sample pair randomly selected from the dataset , Fixed to 1, this is equivalent to uniform sampling on the beta distribution. The interpolated image is then... Input model B and obtain its output. And calculate the sample pair according to the following formula. Mixup loss under model B :

[0057] (3);

[0058] Where n represents the total number of samples.

[0059] This strategy not only implicitly achieves data augmentation but also forces the model to learn the intrinsic semantic features of samples through sample interpolation rather than relying on spurious or noise-related surface cues, thereby effectively improving the model's robustness. This method introduces Mixup loss into the photovoltaic cell defect detection task to enhance the model's resistance to label noise and class imbalance. Furthermore, it is the first to combine Mixup loss with CE loss, constructing a heterogeneous loss training paradigm. Under this heterogeneous loss setting, the two models employ different optimization objectives: Model A uses the CE loss with strong fitting ability to more efficiently capture the overall data distribution; while Model B uses the noise-robust Mixup loss, focusing on learning more stable and generalizable semantic features in the data. The two models complement each other through a subsequent periodic parameter aggregation mechanism, jointly improving the final classification performance.

[0060] Step S32: During heterogeneous loss training, since the two models are optimized independently, there is a lack of effective information interaction between them, making it difficult to fully integrate their respective advantages, thus limiting further improvement in overall performance. Existing noise label learning methods typically use the same loss function for both models and achieve interaction between them by allowing them to select "reliable samples" from each other. However, in photovoltaic cell defect classification tasks, the number of training samples is often limited, and the defect types are diverse and sparsely distributed. If training is based solely on a subset of selected samples, marginal cases or rare defect patterns are easily overlooked, negatively impacting the model's generalization ability. Therefore, a more ideal approach is to allow both models to fully utilize all available samples for learning, rather than relying on a sample selection mechanism to facilitate interaction. To achieve this goal, this invention aggregates the parameters of the two independently optimized models using heterogeneous loss after each number of training epochs. Specifically, the parameters of model A and model B are aggregated and calculated according to the following formula:

[0061] (4);

[0062] in, and These represent the network parameters of model A and model B, respectively. These are the aggregated global parameters.

[0063] To enable the two models to learn from each other and absorb each other's strengths, the aggregation parameter will be reassigned to both models and reinitialized according to the following formula:

[0064] (5)

[0065] Subsequently, the two models continue to be trained collaboratively under heterogeneous loss (i.e., model A uses cross-entropy loss and model B uses Mixup loss), and perform periodic parameter aggregation and re-initialization operations every fixed T epochs (training rounds).

[0066] Through this mechanism, the advantages of different loss functions are preserved: Model A maintains a strong fit to the overall data distribution, while Model B continuously learns semantic features robust to noise and perturbations; simultaneously, implicit interaction and knowledge transfer are achieved through parameter-level fusion. Ultimately, this strategy enables the resulting ensemble model to possess both strong discriminative performance and high robustness to label noise and class imbalance.

[0067] Step S4: Repeat steps S31 and S32 for iterative heterogeneous training until the preset total number of training rounds is reached. ;

[0068] Step S5: Input the electroluminescent image of the photovoltaic cell to be detected into the trained model A and model B, and average the output prediction results of the two models to more effectively utilize the soft label information and obtain the defect classification result of the image to be detected.

[0069] Specifically, for each test image ,in The test set's final prediction result is calculated using the following formula:

[0070] (6);

[0071] in, This indicates the input under two CNN models. The final predicted probability distribution. This strategy helps suppress anomalous predictions that may be generated by a single model, improving prediction stability. Such ensemble methods based on prediction averaging have been widely used in the field of noisy label learning, and this study is the first to introduce them into the photovoltaic cell defect classification task, effectively enhancing the reliability and robustness of the model in real industrial scenarios, such as... Figure 3The image shows a comparison of the proposed method with a baseline method based on cross-entropy on defect images of monocrystalline silicon and polycrystalline silicon, as well as the defect detection results using only model A from this invention. It is evident that this method helps suppress anomalous predictions that may arise from a single model, thereby improving prediction stability.

[0072] Example 2:

[0073] This embodiment provides a dual-network integrated learning method for photovoltaic cell defect detection, including:

[0074] The model building module is used to build and initialize two convolutional neural network models with the same structure, namely model A and model B;

[0075] The preheating training module is used to preheat training the models A and B using the cross-entropy loss function;

[0076] The collaborative training control module, used to control the iterative training process, includes:

[0077] A heterogeneous training unit is used to optimize model A using cross-entropy loss, while simultaneously co-training model B using Mixup loss.

[0078] The parameter aggregation unit is used to periodically aggregate and redistribute the network parameters of model A and model B according to a preset aggregation period T.

[0079] An iterative control module is used to control the collaborative training control module to run repeatedly until the preset total number of training rounds is reached;

[0080] The defect detection module is used to receive the image to be detected, call the trained model A and model B to make predictions, and average the prediction results to output the defect classification result.

[0081] The system executes the following process to implement the method provided in Example 1:

[0082] Model Construction and Initialization: First, construct and initialize two convolutional neural network models with the same structure, denoted as Model A and Model B, respectively. The convolutional neural network can be a mature architecture such as ResNet, VGG, or MobileNet, or a custom network designed for electroluminescent image features.

[0083] Warm-up Training: Subsequently, using the cross-entropy loss function, models A and B are trained on the training dataset for several epochs. The purpose of this stage is to allow the two models to learn the basic feature distribution of the dataset from a randomly initialized state, forming a basic defect discrimination ability, and avoiding the instability that may be caused by directly entering complex co-training from a completely random state. The number of warm-up training epochs can be adjusted according to the size of the dataset. As a preferred but non-restrictive example, 5 training epochs are usually performed. During this process, although the two models are updated independently, they use the same loss function and objective, aiming to converge to a relatively close initial state with initially usable performance.

[0084] Periodic collaborative training: After the warm-up period, the system enters the following state: Figure 2 The illustrated periodic collaborative training loop continues to run until the preset total number of training rounds is reached. In each training round, the following operations are performed:

[0085] Heterogeneous loss calculation and backpropagation:

[0086] For model A, the prediction error is calculated using cross-entropy loss and backpropagation is performed to optimize its parameters.

[0087] For model B, we first perform data augmentation on the input samples using Mixup interpolation based on the Beta distribution, then use Mixup loss to calculate its prediction error and perform backpropagation to optimize its parameters.

[0088] Periodic parameter aggregation and redistribution:

[0089] After each preset aggregation period T (e.g., every 5 training rounds), the system will upload the current network parameters from model A and model B respectively.

[0090] The system performs aggregation operations (such as calculating the arithmetic mean) on the parameters of the uploaded model A and model B to obtain a set of global parameters.

[0091] The system re-downloads the aggregated global parameters and assigns them to Model A and Model B as initial parameters for their next round of collaborative training.

[0092] Once the model training is complete, the system enters the application phase:

[0093] The system receives the electroluminescence (EL) image of the photovoltaic cell to be detected and performs standardized preprocessing, including but not limited to grayscale conversion, size normalization to the network input size, and pixel value normalization, to eliminate imaging differences and ensure that the input data is consistent with the distribution of the training data.

[0094] The preprocessed image is simultaneously input into both trained model A and model B to obtain prediction result 1 and prediction result 2, respectively.

[0095] Ensemble prediction: The prediction results of the two models are averaged, and the category with the highest probability value is selected as the final defect classification result. This ensemble strategy based on soft label averaging can effectively smooth out the random misjudgments that may be generated by a single model, and comprehensively utilize the consensus information of the two models to output more stable and reliable detection conclusions, significantly improving its practicality and credibility in real industrial scenarios.

[0096] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0097] In the embodiments provided in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0098] Furthermore, the functional units in the various embodiments of this application 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. The integrated unit can be implemented in hardware or as a software functional unit.

[0099] If the integrated module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0100] The implementation of all or part of the processes in the methods of the above embodiments can also be accomplished by a computer program product. When the computer program product is run on a terminal device, the terminal device can implement the steps in the various method embodiments described above.

[0101] The embodiments described above are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A dual-network integrated learning method for photovoltaic cell defect detection, characterized in that, The method includes the following steps: Step S1: Construct two convolutional neural network models with identical structures, namely Model A and Model B; Step S2: Perform pre-training on the training dataset using cross-entropy loss for models A and B; Step S3: Perform iterative heterogeneous training on Model A and Model B after preheating training. Each training round includes: Step S31: Continue to optimize model A using cross-entropy loss, and optimize model B using Mixup loss based on sample interpolation. Step S32: Every fixed T training rounds, aggregate the network parameters of model A and model B, and redistribute the aggregated parameters to the two models. Step S4: Repeat steps S31 and S32 for iterative heterogeneous training until the preset total number of training rounds is reached; Step S5: Input the electroluminescent image of the photovoltaic cell to be detected into the trained model A and model B, and average the output prediction results of the two models to obtain the defect classification result of the image to be detected.

2. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 1, characterized in that, In step S2, the preheating training lasts for 5 rounds.

3. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 2, characterized in that, In each round of the preheating training, the formula for the cross-entropy loss function is: ； in This represents the cross-entropy loss, where n represents the total number of samples in a training batch. This represents the image of the i-th sample. Let f represent the hard label of the i-th sample, and let f represent the convolutional neural network model. This represents the network parameters of the model.

4. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 1, characterized in that, For each training sample in the training dataset, the Mixup loss is calculated as follows: Another sample is randomly selected from the training dataset, and a mixing coefficient is sampled from the Beta distribution. The sampled mixing coefficient is used to perform linear interpolation between the image of the current sample and the image of the random sample to generate an interpolated image. The interpolated image is input into the model B, and the cross-entropy loss between its output and the label of the current sample and the label of the random sample is calculated. The Mixup loss for this training sample is obtained by weighting and summing the two cross-entropy losses using the mixing coefficients.

5. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 4, characterized in that, The hyperparameter of the Beta distribution is set to 1.

6. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 1, characterized in that, In step S32, the method for aggregating the network parameters of model A and model B is as follows: ； in, and These represent the network parameters of model A and model B, respectively. These are the aggregated global parameters.

7. The dual-network integrated learning method for photovoltaic cell defect detection according to claim 1, characterized in that, In step S5, averaging the output prediction results of the two models includes: Calculate the arithmetic mean of the output prediction results of Model A and Model B for the photovoltaic cell electroluminescence image to be detected, and select the category with the highest probability value in the average value as the final classification result of the image to be detected.

8. A dual-network integrated learning system for photovoltaic cell defect detection, characterized in that, include: The model building module is used to build and initialize two convolutional neural network models with the same structure, namely model A and model B; The preheating training module is used to preheat training the models A and B using the cross-entropy loss function; The collaborative training control module, used to control the iterative training process, includes: A heterogeneous training unit is used to optimize model A using cross-entropy loss, while simultaneously co-training model B using Mixup loss. The parameter aggregation unit is used to periodically aggregate and redistribute the network parameters of model A and model B according to a preset aggregation period T. An iterative control module is used to control the collaborative training control module to run repeatedly until the preset total number of training rounds is reached; The defect detection module is used to receive the image to be detected, call the trained model A and model B to make predictions, and average the prediction results to output the defect classification result.

Images