Advanced Transfer Learning Technique for Enhanced Detection and Classification of Damaged Solar Cells

The shift from using non-renewable sources of energy to renewable sources of energy is key to the sustainable development of the environment [1, 2]. Renewable energy, energy from natural sources like solar, wind, and hydro energy, has proved to be one of the biggest benefits of green energy as against conventional energy, which usually entails the release of hazardous substances into the environment and exhaustion of resources [3, 4]. Green energy sources help to minimize carbon footprints and global warming. The increased demand and use of renewable energy sources is therefore occasioned by the need to solve environmental concerns and make available energy for future generations [5, 6]. Investing in green energy technologies has been widely encouraged for quite some time now, and this is supported by the fact that as the technology is developed, the costs continue to reduce. For example, sources such as solar energy have come down in price and are more achievable due to enhancements in photovoltaic technologies and materials [7-9].

High-volume solar panel manufacturing and maintenance inspections involve specific assembly processes that require optimal examination techniques [10, 11]. Installing solar panels can, at times, be problematic because of failed parts, as with any other technology or mechanical system as well as Damage to some cells is a common factor that disrupts panel functioning [12]. To address this issue, the next steps typically require a more comprehensive cell examination, which includes a visual inspection to identify any damaged cells. This process can be time-consuming due to the identification of faulty cells and the potential for human error [13, 14]. As a result of the damage found in solar cells, technology has been utilized and developed to create AI, a tool that has the capacity to solve different problems in different fields [15-17]. For the application of AI, a subdomain of machine learning (ML) called deep learning (DL) has shown remarkable results for image classification [18-20]. This makes DL an ideal strategy to handle the detection and inspection of defected solar panels, as described in some studies [21, 22]. Researchers have employed image assessments such as electroluminescence (EL), infrared (IR), and RGB images for the classification of damaged PV cells [23-25]. All these techniques demonstrate that EL inspects PV more efficiently than conventional charge-coupled device (CCD) imaging methodologies [26-28].

Furthermore, authors have confirmed that the use of DL-based algorithms is an intelligent tool for performing many computer vision tasks, including image classification, object detection, recognition, and identifying similarities between two images [29-31].

Several researchers have proposed research propositions for classifying damaged PV cells using a variety of DL paradigms and approaches. For instance, in a work by Chen et al. [32], the authors proposed multispectral CNN networks to classify damaged solar cells based on RGB model images with an accuracy of 94.9%. It offers defect information on an as needed basis but can be expensive and may not be easy to administer.

Similarly, Rahman et al. [33] proposed the use of the U-Net model to identify damaged cells from sequences of EL images, a model that captures details not readily visible in images. The issues are the large data sets and resources required, as well as the fact that these models is computationally intensive. Improvements could be made to the scale of the datasets and the tuning of the models.

In addition, Pierdicca et al. [34] employed transfer learning for the identification of broken and damaged cells using images obtained from the remote sensing system. Still, it demands lots of data and may be uninterpretable. To make the model more comprehensible, future work could focus on data augmentation and explanation.

Cipriani et al. [35] used the VGG16 model and five CNN networks to tell the difference between hotspots and dust objects on PV cells with an accuracy of 98%, even though their model had more layers. The model may not be able to effectively handle other types of defects. Improvements in feature detection and expansion of defect coverage would be the direction of future work.

Akram et al. [36] have proposed a study of distinguishing damaged cells in EL images through a lightweight CNN with an accuracy of 93.02%. The disadvantages of this method include difficulties with data and computation. Here, improvements could be made on points that are related to the use of data and its utilization that do not necessarily require computational processing.

Moreover, Li et al. [37] used DL techniques to perform deep feature extraction for faulty PV cell detection, which may have potential problems associated with a variety of defects. These may include increasing the flexibility in functional learning, which has to do with different types of defects.

On the other hand, Aziz et al. [38] proposed a modification to a pre-trained Alexnet model for faulty detection in PV. Applying it to all types of faults might not yield the same effectiveness. Future work could review and broaden the categories of faults considered

Li et al. [39] applied deep convolutional neural networks to learn fault pattern features to increase diagnosis accuracy. High- and low-tone image quality problems could arise. Some enhancements could be directed toward managing different conditions.

Girshick et al. [40] introduced a paper using Region-CNN combined with SVM for detection operations. It's highly accurate but complex and resource-intensive.

Despite the advancements in image processing and ML algorithms for defect detection, many approaches fail due to the wide variations in defect appearance and environmental changes. Traditional techniques are sometimes insufficient to distinguish similar anomalies, and they may not be applicable to new forms of damage and degradation.

This is the reason that our experiment is designed to redress those limitations with the help of some advanced transfer learning methods and deep convolutional neural networks (CNN). We are utilizing current model, such as InceptionV3, to identify potential optimization techniques that can improve the accuracy of detection percentages and reduce the likelihood of false positives in defect diagnosis. Thus, this approach goes not only beyond the framework of the development of the method of PV module inspection but also beyond the key objectives of improving the efficiency and reliability of energy-green technologies.

The rationale for this study is the quest for better solutions in inspection practices than those that are currently available. By developing and validating a robust model for solar cell damage detection, we hope to support the sustainable growth of solar energy systems and promote the wider adoption of renewable energy solutions.

The following provides a concise summary of the research objectives and its contributions:

(1) Improve the efficiency and accuracy of detecting damaged solar cells in PV panels, with the primary aim of achieving early detection to prevent potential failures and enhance the overall performance of solar panels.

(2) Introduce the concept of transfer learning with the pre-trained InceptionV3 model for classifying damaged versus normal solar cells. This method uses a publicly available dataset of EL images of solar cells and adds to the dataset using data augmentation techniques to make it bigger and more diverse, which improves the generalization of the model.

(3) Show the importance of the presented approach using the illustrative high level of achieved accuracy, which makes it 95.53%. In this paper, we also do comparisons with other earlier-implemented methodologies in order to put into perspective the main features of the method that is developed here as being better and improved. Furthermore, we avoid using a GPU as a parameter in the research and instead use Google Colab to make GPUs available to a wider audience.

The article's classification is as follows: Introduction: This section includes background information, a literature review, the purpose of the study, and its significance. Methodology: This section provides more details about the strategy used in this work, including the methods for transfer learning, dataset, and the planning of the specific experiments. Results and discussions: We analyze the experimental results, compare the proposed method to existing methods, and then discuss the results. Conclusion: Conclusions and future research directions: The study concludes with a concise summary of the results, highlights the study's contributions to the current knowledge base, highlights the study's practical importance in the relevant field, and offers suggestions for future research.

We evaluate the dataset using the pre-trained InceptionV3 model. A 90:10 training-testing ratio has been used, and train and test dataset details are given in Table 2. The augmented data has been used for the training of proposed models

For the train and valid dataset, the images were resized to 224×224 pixels. We maintained the batch size at 10, and trained InceptionV3 for epoch values of 80. We manually selected the batch size value and the number of epochs through empirical means. The learning rate has been fixed to 0.001 for the training of each model, and the Adam optimizer has been utilized for error minimization purposes. Furthermore, we measured the performance of the InceptionV3 model using metrics like specificity (Spe), sensitivity (Sen), precision (Pre), F1-score, and accuracy (Acc). These metrics were obtained by various parameters of the confusion matrix, such as true positive (TP), true negative (TN), false positive (FP), and false negative (FN) [48]. These metrics have been calculated using Eqs. (1)-(6).

$precision =\frac{T P}{T P+F P}$     (1)

$Recall =\frac{T P}{T P+F N}$     (2)

$sensitivity =\frac{T P}{F N+T P}$     (3)

${s p e c i f i c i t y}=\frac{T N}{F P+T N}$     (4)

${ A c c u r a c y }=\frac{T P+T N}{F P+T P+T N+F N}$     (5)

${F 1}\ score =2 \times \frac{{ Precision } \times { Recall }}{ { Precision }+{ Recall }}$     (6)

In this study, normal and defective were considered as negative and positive classes, respectively. Thus, the TN and TP represent accurately predicted normal and defective classes. Whereas FN and FP indicate misclassified predicted normal cases and defective cases, respectively.

Furthermore, the performance of InceptionV3 has been compared in terms of parameters such as training loss, validation loss, and validation accuracy at each epoch value. Table 3 presents the results of these parameters. We measured the parameters to identify instances of over-fitting and under-fitting in the trained models. Figure 4 displays the graphs of training loss versus validation loss for each model under study.

For further performance validation, confusion matrices have been generated to classify true positive, true negative, false positive, and false negative values after training. Figure 5 displays the confusion matrices from the test dataset and the loss graphs from the train and valid datasets of the InceptionV3 model.

Table 2. Data splitting details

Table 3. Training performance of InceptionV3 model

214.png

Figure 4. Loss graph for train and valid dataset for the InceptionV3 model

215.png

Figure 5. Confusion matrix on test dataset for the InceptionV3 model

Table 4. Performance of InceptionV3 model

Eqs. (1)-(6) have evaluated the performance parameters of the trained models, such as precision, recall, F1 score, sensitivity, specificity, and accuracy, from the numbers obtained through confusion matrices. Table 4 displays the results of these parameters.

3.1 Comparison with different optimizers

Four different optimizers, such as Root Mean Square Propagation (RMSprop), An Adaptive Learning Rate Method (Adadelta), Stochastic gradient descent (SGD), and Adam, which combines the benefits of both Adaptive Moment Estimation (Adam) were used to compare the performance on Iceptionv3 models and evaluation of the optimizer and to select the optimizer that could be the best optimizer applied to the model. The Adam optimizer exhibited promising results among all other optimizers. Therefore, we selected Adam Optimizer for the model's training. The results in Table 5 were calculated using the Adam optimizer.

Table 5. Performance of InceptionV3 model

3.2 Comparison with different batch size

Batch size is considered to be one of the important hyperparameters for deep neural networks. This research provides information on the impact of different batch sizes. Table 6 outlines the accuracy of tests conducted on various batch sizes, including 8, 10, and 12. We have observed that a batch size of 10 yields better testing performance. Thus, a batch size of 10 has been kept training the InceptionV3 model.

Table 6. Evaluation of accuracy for varying batch sizes using the InceptionV3 model

3.3 Comparison with other methods

Joshua et al. [49] designed an intelligent method used for detecting the fault in the solar panel at Kangwon National University, Samcheok Campus. By using the deep learning model ResNet-50, the goal of the paper is to find errors and improve the accuracy of data from the university's solar-hydrogen system through data preprocessing. Metrics, general accuracy, and loss specifically indicate its potential for real-time monitoring and maintenance.

However, there are several drawbacks as well: first, the resultant model is less able to detect many sources of fault; second, it also has a restricted variety of faults that are detected, and the used dataset is taken only from a single campus. Painstaking measures such as cross-validation could be implemented, while metrics such as precision and recall could be useful in giving a broader view of the model's evaluation, all in a bid to reduce cases of the model being overfit and thereby underperforming in new data.

Moreover, Kaur et al. [50] used VGG16 in addressing the task of fault detection in PV panels. The model is trained on 885 images across 15 epochs and can correctly classify different faults into the six classes with 86% accuracy. Despite the potential for real-world implementation, there are limitations to the work: the small dataset that we applied the solution to might impact generalization; the machine learning model was trained on a short time span, and while results show that it outperforms other techniques on unseen data after adapting it to a new training set, it remains limited in terms of the time that it has been exposed to.

Additionally, the absence of performance measures like precision, recall, and F1-score for some steps complicates the evaluation of the methods, particularly when dealing with imbalanced datasets. Another adjustment that could potentially yield better results is the change in the learning rate.

Furthermore, Su et al. [51] looked into how supervised learning using CNN and the addition of thermographic images could help classify defects in PV modules. Their CNN achieved 92.5% accuracy for anomaly identification and 78.85% accuracy in defect classification in eight classes. This method is efficient in processing huge amounts of data and detecting outliers, but it is also obstructed by large within-class and between-class variations that decrease the efficiency of defect classification, especially in the case of an unbalanced dataset.

Furthermore, to classify PV panels into different classes, namely healthy, non-faulty hotspot, and faulty, Ali et al. [52] used a fusion of SVM models by using RGB, texture, HOG, and LBP features. Their model demonstrated 92% testing accuracy. It is considered to have the fewest computational requirements and a greater a greater need for storage than any other ML algorithm. However, it relies on specific feature sets that cannot be expanded to encompass other types of defects, and it necessitates a significant number of tests to ultimately identify the optimal features for collaboration [53].

Moreover, Zhang and Yin [54] proposed an improved YOLO v5 model that can solve issues, including complex backgrounds and irregular and differently sized defects in solar cells. The method utilizes deformable convolution, an ECA-Net attention mechanism, and a tiny defect prediction head, resulting in an 89.64%. It has a comparatively low accuracy and a speed of 36.24 FPS. It performs very well in online detection and produces relatively accurate results at different scales. It may have some issues when applied to datasets with different characteristics, and it is computationally intensive.

Therefore, in our paper, InceptionV3 Method was selected because it has better feature extraction and classification performances, with 95.53%. The discrimination accuracy of detecting defects. That is why, when it comes to dealing with defects of different sizes and different characteristics, multi-scale feature extraction in InceptionV3 turns out to be of great use. This method also achieves a balance between error and efficiency, making it the most suitable for the study goals.

3.4 Validate the model using an external dataset

In order to test the generalizability and the solidity of the proposed model, we performed the validation on an external data set. In particular, we work with the “Dust Detection on Solar Panel Using InceptionV3” [55] the dataset presented on Kaggle.ally, we tested the model on the "Dust Detection on Solar Panel Using InceptionV3" by Afroz in the dataset available on Kaggle. This dataset entails images of solar power in diverse states, which include those with dust on the panel and those that don’t. In order to achieve this, we utilized an external dataset, distinct from the training data, for our evaluation. This allowed us to assess the generality of the obtained model and its performance on new data, both of which are crucial for a practical solar panel defect detection system. Our method adopts the transfer learning strategy, using the InceptionV3 model as a basis for refinement.

 In this work, the InceptionV3 convolutional neural network that was trained on the ImageNet database was fine-tuned to specifically perform the task of identifying defects in solar cells. By fine-tuning, we refine the final layers of the pre-trained model to identify characteristics associated with electroluminescence images of solar cells. This transfer learning technique enhances performance by learning features in limited datasets such as ImageNet [56, 57].

For the external validation, we made sure that the images from the Kaggle dust detection dataset were prepossessed in the same manner as used in the original dataset. We also applied certain data pre-processing techniques, which include resizing the images to 300x300 pixels, scaling the obtained pixel intensities to the range of 0 and 1 by dividing them by 255, and applying data augmentation techniques such as random rotation by up to 30 degrees and horizontal and vertical flipping of the images. These steps were followed in order to avoid influencing and having a bias in the validation process.

To assess the model’s performance on the external dataset, basic measures of precision, recall, accuracy, and F1 score values were used. It was shown that when the model foresight its performance on the external Cavin dataset, it achieves rather good values of the metrics that were observed during training and the first tests. This outside opinion supports the idea that the model can successfully generalize to unseen data, indicating that utilizing the transfer learning approach to detect defects on solar panels is viable and still feasible regardless of the different circumstances a solar panel might be under.