Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.
Advertisement
Scientific Reports volume 16, Article number: 21371 (2026) Cite this article
1531 Accesses
Metrics details
As the global transition toward renewable energy accelerates, ensuring the operational reliability of photovoltaic (PV) systems has become increasingly critical. Manual inspection procedures remain labor-intensive and economically inefficient, particularly in large-scale solar installations. This study proposes an automated diagnostic framework for classifying anomalies associated with PV panels using deep learning approaches built on the EfficientNet as a backbone, techniques such as Fine-tuned Transfer Learning (FTL), Deep Feature Extraction + Classifier (DFE-C), and Alternative Fine-tuning Setup (AFS) are applied and their efficacy in detecting various defect categories is evaluated. Furthermore, the analysis focuses on a comprehensive evaluation of these strategies in terms of robustness and accuracy regarding classification capabilities. The approaches are evaluated under a consistent data partitioning strategy derived from the same PV image dataset, enabling a systematic comparison of their classification accuracy, robustness, and consistency. The results indicate that FTL enhances domain adaptability, while DFE-C exhibits the greatest overall stability and performance under limited and variable data conditions. The AFS approach provides a balanced trade-off between flexibility and convergence. The experimental framework incorporates structured training pipelines, hyperparameter control, and performance benchmarking using accuracy, macro F1-score, and fold-based stability analysis. Specifically, the DFE-C approach achieved a superior overall accuracy of 94.05%, demonstrating near-perfect diagnostic capability in critical categories such as physical damage (100%) and snow coverage (96%). In short, the proposed framework provides a comparative evaluation methodology for automated PV inspection and offers useful insights for the development of more reliable AI-based diagnostic systems for PV energy applications.
The global energy landscape is undergoing a profound transformation toward sustainable and renewable sources to mitigate the impacts of climate change and achieve long-term carbon neutrality. Within this transition, solar PV energy has emerged as one of the most promising technologies due to its scalability, rapidly decreasing installation costs, and minimal environmental impact during operation. Nevertheless, the long-term economic viability and energy efficiency of PV systems are strongly dependent on their operational condition and maintenance effectiveness. Exposed continuously to harsh outdoor environments, PV modules are prone to a wide range of defects, including micro-cracks, hotspots, snail trails, and surface contamination. If such anomalies remain undetected, they can result in substantial energy yield losses, accelerated material degradation, and, in severe cases, safety risks such as electrical faults or fire hazards. Advanced data augmentation techniques, specifically those utilizing GANs for infrared and visible image pairs, have been proposed to overcome small-sample constraints and improve fine-grained fault recognition¹.
Conventional O&M practices for large-scale solar installations primarily depend on manual visual inspections and periodic thermographic surveys performed by human operators. While these methods are widely adopted, they are inherently labor-intensive, time-consuming, and susceptible to subjective interpretation and human error. As utility-scale PV plants expand to megawatt- and gigawatt-level capacities, traditional inspection approaches become increasingly inefficient and economically unsustainable. Consequently, there is a growing demand for automated, scalable, and reliable inspection solutions that can support proactive maintenance strategies.
Recent advances in AI and computer vision have driven the emergence of a new generation of automated PV inspection systems. In particular, deep learning frameworks like CNNs and VTs have demonstrated robust capabilities in detecting and classifying PV panel defects using both RGB and infrared imagery. By enabling rapid, non-contact, and data-driven fault identification with high accuracy, these AI-based approaches offer significant advantages over conventional techniques. Recent research has introduced multimodal feature fusion models, such as RFE-YOLO, which differentiate feature enhancement for visible and infrared modalities to achieve high-precision results in centralized power plants².
Despite the rapid growth of AI-based diagnostic approaches for PV maintenance, a critical challenge in the existing literature is the need for more rigorous evaluation of model robustness and consistency. Most prior studies assess model performance using a single dataset, which may lead to overly optimistic results due to dataset-specific biases. In such cases, high accuracy values may not fully reflect the behavior of a model across different data partitions or varying visual patterns within the same dataset. This highlights the importance of comparative evaluations under consistent experimental settings. Furthermore, hierarchical deep learning frameworks integrated with dataset-balancing strategies are being developed to address the inherent class imbalance and low thermal contrast in large-scale PV inspection datasets³.
In response to this challenge, this paper presents a comparative benchmarking of three distinct AI architectures for automated PV panel defect detection and maintenance support. The study is conducted on a publicly available PV image dataset and focuses on a controlled comparison of different learning strategies under a unified experimental setting.
This work does not introduce a new network architecture. Instead, its novelty lies in the systematic comparison of different transfer learning strategies for multi class PV defect classification. By evaluating fine tuning and feature based methods under a consistent experimental setup, we provide a more detailed analysis of model performance in this specific field. The main contributions of this study can be summarized as follows:
A comparative performance evaluation of three state-of-the-art AI models for image-based PV defect detection.
An in-depth analysis of model consistency and robustness under a unified experimental setting, including train/validation/test evaluation and fold-based validation for all compared methods.
A quantitative comparison of classification accuracy and computational efficiency, providing practical insights for selecting suitable AI models in PV maintenance applications.
The remainder of this paper is organized as follows. Section 2 presents a review of related work in AI-based PV inspection. Section 3 describes the problem formulation and the dataset employed. Section 4 details the proposed methodology and AI architecture. Section 5 reports and discusses the experimental results, and Section 6 concludes the paper with key findings and directions for future research.
Despite the progress achieved in AI-based PV panel inspection, several challenges still remain for the broader use of image-based diagnostic systems in maintenance practice. In this context, the main issues can be grouped into three categories:
The Generalization Gap and Dataset Dependency: Most existing AI models for PV fault detection are developed using localized datasets collected under specific environmental and imaging conditions. This practice leads to “dataset-dependent learning,” where models overfit to specific features such as a particular camera’s sensor noise, specific solar irradiance levels, or unique geographic backgrounds. Consequently, models that show high accuracy on training data often suffer from significant performance degradation when deployed on unseen datasets with different environmental characteristics. To bridge this gap, recent frameworks have utilized cross-domain adaptive generative adversarial networks to align feature distributions across disparate PV installations, effectively reducing domain-specific biases⁴. There is still a need for more research on model robustness and consistency, which is important for developing reliable PV maintenance tools.
Absence of Systematic Architectural Benchmarking: The current literature frequently proposes individual AI models in isolation, without rigorous, head-to-head comparisons under identical experimental conditions. This lack of comparative evaluation makes it difficult for O&M providers to determine the most suitable architecture for specific maintenance tasks among various options such as standard CNNs, object detection models, or Vision Transformers. Recent benchmarking studies using novel datasets have demonstrated that while newer object detection frameworks offer superior precision, they exhibit distinct trade-offs in Sensitivity when compared to custom convolutional architectures for identifying specific defects like physical damage versus surface contamination⁵ Without a multi-model benchmarking framework, the industry lacks clear guidance on the trade-offs between different AI architectures in terms of accuracy and reliability.
Misalignment Between Accuracy and Operational Feasibility: Academic research often prioritizes raw classification accuracy while overlooking practical maintenance constraints In practical PV inspection settings, factors such as computational efficiency, inference latency, and the relevance of fault types (e.g., distinguishing between temporary soiling and permanent physical damage) can be as important as accuracy. Many models are too computationally expensive for edge-device deployment (like drones) or fail to prioritize defects based on their actual impact on energy yield. The emergence of lightweight models, such as those based on MobileNetV3 and EfficientNet, has shown that it is possible to maintain over 98% diagnostic accuracy while meeting the strict memory and latency requirements of real-time aerial inspections⁶. A broader evaluation perspective should also consider the relationship between AI-based performance metrics and practical maintenance requirements.
The integration of Artificial Intelligence into the operation and maintenance of PV systems has witnessed substantial growth over the last decade, transitioning from manual feature extraction to sophisticated deep learning frameworks. Early research in this domain primarily focused on traditional image processing techniques and shallow machine learning models; however, the recent surge in computational power and the availability of large-scale datasets have shifted the academic focus toward CNNs and, more recently, VTs. While literature is rich with studies proposing novel architectures for defect classification and localization, there remains a critical need to evaluate these models under diverse environmental conditions to ensure their practical reliability.
The following survey of literature examines recent advancements in automated PV inspection, categorized by architectural evolution and evaluation methodologies. These studies highlight the shift from localized, single-dataset assessments toward more complex, multi-model frameworks. By analyzing the strengths and limitations of existing approaches, this section establishes the foundation for our comparative analysis and justifies the necessity of cross-dataset validation in developing robust maintenance-oriented AI tools.
The maintenance of PV systems is critical as surface soiling and structural damage significantly reduce energy efficiency and grid integration reliability⁷. Recent literature has therefore explored automated PV fault diagnosis using hierarchical hybrid architectures. For example, the study in⁷ combined a vision transformer architecture with lightweight CNNs, where a fine-tuned DINOv2 model was used for high-level categorization and enhanced EfficientNetB0 models were employed to improve specific classification accuracy through knowledge distillation and attention mechanisms. The reported results reached up to 98.73% accuracy for damage detection and 93.33% for soiling classification. These findings indicate the potential of hybrid multi-stage models for improving PV fault classification performance.
Intelligent fault monitoring systems play a critical role in enhancing efficiency and reducing maintenance costs in solar energy production. This study compares five machine learning algorithms, including SVM, ANN, and Random Forest, to detect faults caused by partial shading and soiling⁸. Analysis using real-world data from two distinct PV systems demonstrates that models can identify faults using electrical and environmental data. A key finding is that models trained on one system experience significant performance degradation when applied to another system with different characteristics. The ANN model was reported as the most successful, achieving over 98% precision. This research emphasizes that including data from every new installation in the training phase is essential for achieving model generalization.
The accurate detection of surface defects on PV panels is vital for maintaining energy efficiency and minimizing repair expenses. This review paper systematically examines the evolution of the YOLO architecture from v1 to v10 and its contributions to PV quality control⁹. Specifically, the advantages of the single-stage object detection approach for operational efficiency are highlighted. The study notes that while current literature focuses heavily on micro-crack detection, future research should address a broader range of faults. Furthermore, the importance of integrating attention mechanisms into the YOLO architecture to detect complex and small-scale defects is discussed. The article anticipates that future research will advance toward edge computing and federated learning solutions.
Image analysis techniques used for assessing the condition of solar power plants have evolved significantly during 2024 and 2025. This comprehensive overview summarizes the most recent CNN-based research concerning PV modules and power plants¹⁰. The study emphasizes the complementary roles of IR and RGB imaging modes in identifying various fault sources. Infrared (IR) imaging reveals electrical heating-related faults, while RGB imaging is effective for mechanical defects such as glass breakage and discoloration. CNN models have been proven to increase operational speed and reduce downtime in real-time detection systems. The article highlights that automated inspection systems have become an indispensable necessity alongside the increasing volume of panel production.
A new deep learning-based pipeline is proposed for the automated detection of soiling, snow accumulation, and physical damage on solar panels. This research utilizes the YOLOv11 model on a heterogeneous dataset of approximately 9,000 images to identify five different anomaly classes¹¹. The developed model achieved over 95% accuracy and recall rates under varying light conditions and different soiling scenarios. An interactive dashboard created as part of the study processes live camera streams to generate real-time alerts and support proactive maintenance planning. Results indicate that modern architectures like YOLOv11 are superior to traditional manual inspections in both speed and precision. This approach offers a low-cost and scalable maintenance solution for large-scale solar farms.
Recent studies in literature have extensively investigated deep learning techniques for defect detection in PV modules using visual data¹². A comprehensive review analyzes RGB, thermal, and electroluminescence imaging modalities and categorizes existing approaches based on network architectures and fault types. The study highlights that convolutional neural networks dominate current research due to their strong feature extraction capability. However, most reported results are obtained from single-dataset experiments. Consequently, the lack of cross-dataset validation is identified as a major limitation for real-world deployment.
In an effort to enable real-time inspection, a lightweight deep learning framework has been introduced for fault detection in PV modules using thermal imagery¹³. The proposed model focuses on reducing computational complexity while preserving classification accuracy. Experimental results demonstrate that efficient architecture can achieve performance comparable to deeper networks under controlled conditions. Nevertheless, training and testing are conducted on a single dataset. The authors note that robustness under varying environmental conditions remains an open challenge.
Thermography-based fault detection has also been explored for both monofacial and bifacial PV modules, revealing distinct thermal characteristics across module technologies¹⁴. By applying convolutional neural networks to infrared images, the study demonstrates high classification accuracy within the evaluated dataset. Differences in thermal behavior are shown to influence model performance. However, the analysis is limited to data from a single source. As a result, generalization across heterogeneous datasets is not addressed.
Hybrid optimization-based deep learning approaches have also been explored for PV panel fault detection. For example, the workflow shown in Fig. 1 corresponds to the method reported in¹⁵, where a convolutional neural network was combined with bacterial foraging optimization (BFO) to improve feature selection and classification performance. In that study, the processing pipeline included image acquisition, preprocessing, feature extraction, optimization, and final fault classification. The reported results indicated improved accuracy over the baseline CNN on the same dataset. This example shows that optimization-assisted feature selection has also been investigated in the PV fault diagnosis literature.
Example of a BFO-assisted PV fault detection framework reported in¹⁵.
Fault detection in solar energy systems has further been addressed through deep learning-based image analysis frameworks¹⁶. These approaches demonstrate that CNN architecture can effectively identify abnormal operating conditions with high classification accuracy. Standard performance metrics are used to validate the proposed models. However, both training and testing data originate from the same distribution. Domain shift and cross-dataset performance are not considered.
Infrared image-based classification of PV module faults has been investigated using supervised deep learning techniques¹⁷. The study focuses on distinguishing different thermal fault patterns under controlled conditions. Results show strong performance and resilience to noise within the dataset. Nevertheless, the evaluation does not extend beyond a single infrared dataset. The authors emphasize the need for broader validation to ensure real-world applicability.
Large-scale PV array inspection using infrared computer vision has been explored with a focus on utility-scale deployments¹⁸. By integrating thermal imaging with deep learning-based anomaly detection, the study demonstrates effective fault localization across extensive PV installations. The results highlight the feasibility of automated inspection for large solar farms. However, experiments are conducted under specific operational conditions. The absence of standardized multi-dataset benchmarking limits conclusions on generalization.
Transfer learning-based strategies have also shown promising results in PV fault diagnosis. Korkmaz and Acikgoz¹⁹ proposed a multi-scale CNN framework for thermographic PV fault classification, where multiple convolutional branches were designed to capture features at different visual scales while benefiting from pre-trained knowledge through transfer learning. To address class imbalance, the study additionally employed offline augmentation with oversampling, which contributed to improved representation of minority classes. Using 11 different PV fault categories, their method achieved high classification accuracy and demonstrated that transfer learning can provide robust and effective performance for multi-class PV fault detection.
More recently, EfficientNet-based architectures have also been explored for photovoltaic defect analysis. For example, Liu et al²⁰. proposed a lightweight PV cell defect detector based on EfficientNet-B0 for electroluminescence image analysis. In their study, CLAHE-based contrast enhancement was applied to improve the visibility of defect patterns, and a graph channel attention module was introduced to enrich global feature representation beyond local receptive fields. The model was further trained with focal loss to improve the detection of challenging defect categories, and the reported results demonstrated strong performance on the PVEL dataset. This study indicates that EfficientNet-based backbones can provide an effective balance between diagnostic accuracy and computational efficiency in PV defect detection.
Hybrid deep-feature and classical classifier pipelines have also been explored for PV fault diagnosis. Venkatesh et al²¹. used a pre-trained AlexNet model to extract visual features from aerial images of PV modules and then applied feature selection before training multiple conventional classifiers, including SVM, kNN, random forest, and multilayer perceptron. Based on the best-performing base learners, they further constructed stacking ensemble groups and reported very high classification accuracy. Their findings suggest that combining deep feature extraction with classical machine learning classifiers can be an effective strategy for visible PV fault classification.
Despite the promising results reported in recent studies, controlled comparisons between different learning strategies for multi-class PV fault classification remain limited. In particular, the relative behavior of fine-tuned transfer learning, deep feature extraction with a classical classifier, and alternative fine-tuning configurations has not been sufficiently examined under a unified evaluation framework. Therefore, this study aims to provide a comparative assessment of these strategies within a common five-fold cross-validation setting.
A summary of the reviewed studies, including their problem definitions, employed models, evaluation metrics, and reported performances, is presented in Table 1.
The methodology of this research focuses on the application of deep learning techniques to identify surface defects in PV panels which directly affect energy production efficiency. This section describes the systematic approach adopted in this study, including the characteristics of the dataset and the various stages of the experimental process. It outlines the data preparation procedures, which are critical for achieving robust PV defect detection performance, and presents the implementation of the deep learning architecture used to obtain reliable classification results²².The overall procedure is structured to maintain consistency and allow for the evaluation of the model performance in a controlled environment while ensuring the reproducibility of the findings.
The dataset used in this study is a publicly available Kaggle RGB image dataset for solar PV panel fault classification²³. It contains six classes, namely Bird-drop, Clean, Dusty, Electrical-damage, Physical-Damage, and Snow-Covered. A detailed summary of the dataset, including the total number of images and class-wise distribution, is provided in Table 2.
Since the original dataset does not provide detailed metadata on acquisition conditions, sensor settings, or source diversity, these aspects could not be analyzed further in the present study. To support visual interpretation, representative sample images from each class are presented in Fig. 2.
Sample images of different solar panel defect classes.
In addition, a statistical overview of the number of samples in each category is provided in Fig. 3. Examining the class distribution before model training is important to reduce the risk of bias toward a dominant category and to better understand the data structure used in the study. This visual and statistical overview provides a clear basis for the subsequent preprocessing and model evaluation stages.
Distribution of images across the six defect classes.
We also analyzed the numerical distribution of the images across the different classes to ensure that the training process would not be biased towards a single category. Assessing class distribution prior to model training is essential to mitigate performance degradation caused by class imbalance, which has been widely discussed in image-based PV fault classification studies²⁴. This visual and statistical preparation provides a clear foundation for the subsequent pre-processing and model training stages of the study.
Before initiating the training phase, a series of data preprocessing steps were applied to ensure compatibility between the raw solar panel images and the adopted deep learning architecture. In this study, the EfficientNetB0 model was employed; therefore, all images were resized to a fixed resolution of 224 × 224 pixels, which corresponds to the standard input size of the network. This resizing step guarantees consistent spatial dimensions across the dataset and enables stable feature extraction through the convolutional layers, consistent with the architectural design principles of EfficientNet-based models²⁵.
Following resizing, pixel-level normalization was performed using the model-specific preprocessing function associated with EfficientNet. This normalization strategy scales the input pixel values according to the statistical distribution used during the original ImageNet pre-training phase, thereby ensuring alignment with the pre-trained weights. Such normalization improves numerical stability during optimization, accelerates convergence, and enhances the effectiveness of transfer learning.
To assess the generalization performance of the proposed framework and to control overfitting, an initial train–validation–test split was employed for model development and comparative evaluation. After identifying the best-performing approach (DFE-C), a 5-fold cross-validation strategy was subsequently applied to all compared approaches in order to obtain a more robust and statistically reliable performance assessment. The application of k-fold cross-validation has been widely adopted in recent deep learning and engineering studies to obtain more reliable generalization estimates and reduce performance variance caused by a single data partition²⁶.
In the 5-fold cross-validation process, the entire dataset was partitioned into five mutually exclusive subsets of approximately equal size. In each iteration, one subset was used as the test set, while the remaining subsets were utilized for training and validation. This process was repeated five times so that each subset served as the test set exactly once. The overall workflow and data partitioning strategy of the 5-fold cross-validation applied in this study are illustrated in Fig. 4.
Schematic illustration of the 5-fold cross-validation procedure applied in this study.
To enhance the robustness of the proposed classification framework and to improve the model’s ability to generalize under varying image conditions, data augmentation techniques were applied during the training phase. In PV inspection applications, solar panel images may be captured under varying environmental conditions, viewing angles, and illumination levels. Therefore, exposing the model to such variations during training is essential for reliable defect detection, as data augmentation has been shown to significantly improve generalization performance in deep learning-based image classification tasks²⁷.
In this study, data augmentation was applied only to the training images, while validation and test sets were kept unchanged to ensure unbiased performance evaluation. The applied augmentation operations included random rotations, zooming, and spatial translations, which simulate variations in camera positioning and panel alignment. In addition, horizontal and vertical flipping were employed, as surface-level faults on PV panels do not exhibit a fixed orientation and may appear in different directions depending on installation geometry.
The primary objective of these augmentation strategies was to reduce overfitting by preventing the model from memorizing specific visual patterns in the training data, thereby encouraging the learning of more general and discriminative features. This approach effectively increases the diversity of the training samples without requiring additional physical data collection from the field.
As illustrated in Fig. 5, various geometric and spatial transformations were utilized to simulate diverse environmental conditions and capture-angle variations, ensuring the model’s robustness across different solar panel orientations.
Representative examples of the data augmentation techniques applied to the PV panel dataset.
To identify the most effective architecture for image-based solar PV fault detection, a comparative evaluation was conducted using three deep learning–based methodologies built upon the EfficientNetB0 backbone. Rather than comparing unrelated network families, this study focuses on analyzing how different learning paradigms applied to the same feature extractor influence classification performance, robustness, and generalization capability, an evaluation strategy frequently adopted in recent transfer learning studies to ensure controlled backbone comparison²⁸. The evaluated deep learning approaches and their general characteristics are summarized in Table 3.
The first approach FTL follows a fine-tuned transfer learning strategy, where the EfficientNetB0 model pre-trained on the ImageNet dataset is employed with a custom classification head.
Let (:mathcal{x}in::{mathbb{R}}^{224times:224times:3})denote an input image and (:{Phi:}left(mathcal{x};{theta:}_{pre}right))represent the feature extraction function of the pre-trained EfficientNetB0 backbone. In the fine-tuning configuration, the parameter set is partially updated such that(::theta:=:{theta:}_{pre}^{*}cup:{theta:}_{new}), where selected higher-level layers and the newly added classifier weights are optimized during training. This strategy enables end-to-end learning while preserving generic low-level representations and has been shown to improve domain adaptation performance in recent deep transfer learning frameworks²⁹.
The second approach DFE-C adopts a feature extraction–based learning paradigm. In this approach, deep feature embeddings are obtained as shown in Eq. (1).
where (:mathcal{x}) denotes the input RGB PV image, (:{Phi:}left(.;{theta:}_{pre}right)) represents the pre-trained EfficientNetB0 feature extractor with frozen parameters (:{theta:}_{pre}), and (:f) is the extracted deep feature vector. These fixed high-level representations are subsequently provided to a conventional classifier to produce the final prediction.
By decoupling feature learning from the classification stage, this approach allows for a more flexible decision boundary and facilitates robust performance analysis, particularly under limited data conditions.
In the present implementation, the extracted deep feature vectors were standardized before classification, and a Linear SVM was used as the final classifier. The SVM regularization parameter was selected during model development on the training data of each fold. A linear SVM was preferred to maintain a simple and stable hybrid classification pipeline, while keeping the focus of the study on comparing different learning strategies rather than exhaustively benchmarking multiple conventional classifiers.
Both approaches leverage the same pre-trained EfficientNetB0 architecture and identical input preprocessing steps, ensuring a fair and controlled comparison. The observed performance differences can therefore be attributed primarily to the learning strategy rather than architectural discrepancies.
The third approach, AFS, adopts an alternative fine-tuning strategy based on EfficientNetB3. Unlike DFE-C, which uses fixed deep features and a separate classifier, AFS performs end-to-end learning by adapting the higher-level layers of the backbone to the PV fault classification task. This configuration allows the model to learn more task-specific representations, while also introducing greater sensitivity to optimization and fold-to-fold variation.
The structural distinctions between the three evaluated approaches are schematically illustrated in Fig. 6.
Architectural schemes of the three evaluated methodologies.
To reduce the effect of class imbalance, class weights were applied during training for the end-to-end deep learning approaches, namely FTL and AFS. For each training split, the class weight for class (:i) was computed as(:::{w}_{i}=N/:K.{n}_{i}:) where (::N:)is the total number of training samples, (:K:) is the number of classes, and (:{n}_{i}) is the number of samples in class (:i). In the five-fold setting, these weights were computed separately for the training portion of each fold and applied during model fitting. The DFE-C approach followed a separate feature extraction and SVM-based classification pipeline.
The training and evaluation framework was designed to objectively compare the strengths and limitations of each approach in terms of classification accuracy, generalization capability, and robustness to surface-level variations commonly observed in field-acquired solar panel images. Based on this comparative analysis, all approaches were further evaluated using a 5-fold cross-validation scheme, as discussed in the subsequent sections.
The experimental setup was designed to ensure a fair and reproducible comparison among the evaluated approaches. All input images were processed according to the preprocessing steps described in Sect. 4.2, and training procedures were defined based strictly on the adopted learning paradigm of each approach. All experiments in this study were conducted using training, validation, and test splits derived from the same dataset. No external dataset was used in this study.
For the deep learning–based approaches employing transfer learning, EfficientNetB0 and EfficientNetB3 models were trained using a batch size of 32. A two-stage training strategy was applied, consisting of an initial head-only training phase followed by partial fine-tuning of higher-level convolutional layers. Each stage was trained for a maximum of 20 epochs, resulting in an upper bound of 40 epochs per approach. Such staged fine-tuning strategies have recently been reported to improve stability and domain adaptation performance in vision-based transfer learning frameworks³⁰. To prevent overfitting, early stopping was applied in all deep learning approaches, monitoring validation accuracy with a patience value of 6, which typically led to convergence between 12 and 18 epochs.
The Adam optimizer was used for all deep learning–based approaches. The learning rate was gradually reduced from 1 × 10^{− 3} to 1 × 10^{− 4} during fine-tuning. Given the multi-class nature of the solar PV defect dataset, weighted categorical cross-entropy was adopted as the loss function³¹, defined in Eq. (2).
where (:N) is the number of training samples, (:C) is the number of classes, (:{y}_{ic}in:) {0,1} denotes the one-hot encoded ground-truth label of sample (:i) for class (:c), (:{p}_{ic})is the predicted probability for class (:c) obtained via the SoftMax function³¹, and (:{w}_{c}) is the class weight assigned to class (:c).
The use of cross-entropy loss in conjunction with adaptive optimization strategies has been widely adopted in modern deep neural network training frameworks for stable convergence and improved generalization²⁶. In addition, class weighting was applied during training to mitigate the effects of class imbalance and improve recognition performance for underrepresented defect categories.
In contrast, the feature-based configuration utilizing EfficientNetB0 embeddings combined with a Linear SVM classifier did not involve epoch-based deep model training. In this case, the EfficientNetB0 network was used solely as a fixed feature extractor with frozen weights, generating embeddings using the feature extraction function defined in Eq. (1).
Feature extraction was performed with a batch size of 64. The subsequent classification was performed using a Linear SVM as the final classifier. The training dynamics, including the two-stage strategy for fine-tuned models and the static extraction process for the SVM-based configuration, are central to the experimental design. A summary of the primary training parameters is provided in Table 4, while the corresponding procedural workflows for each learning paradigm are schematically illustrated in Fig. 7.
Experimental workflow of the training and feature extraction phases.
This section presents the experimental results obtained from the evaluation of the approaches described in Sect. 4. The classification performance of the models was quantified using standard multi-class evaluation metrics, and the results are reported in both tabular and graphical forms.
The overall performance metrics, including accuracy, precision, recall, and F1-score, are summarized in Table 5. All compared methods were evaluated using a common 5-fold cross-validation framework, and the results are reported as mean ± standard deviation.
According to Table 5, AFS yielded the highest mean accuracy, DFE-C provided the best balance in terms of macro F1-score and stability, and FTL remained the weakest of the three compared approaches. Overall, all three methods achieved strong multi-class classification performance on the six-class solar PV fault dataset, although clear differences were observed in terms of average performance and fold-to-fold stability. FTL, based on fine-tuned EfficientNetB0, produced the lowest overall results among the compared methods, with a mean accuracy of 92.66% and a mean macro F1-score of 92.84%. While this method remained competitive, its performance showed greater sensitivity to data partitioning than the stronger alternatives. DFE-C, which combines EfficientNetB0-based feature extraction with a Linear SVM classifier, achieved a mean accuracy of 94.05% and the highest mean macro F1-score of 94.33%. It also exhibited the lowest standard deviation values among the compared approaches, indicating the most stable and consistent behavior across folds. AFS, based on fine-tuned EfficientNetB3, achieved the highest mean accuracy at 94.18%, confirming the strong representational capability of the deeper backbone architecture. However, its larger standard deviation values indicate greater variability across folds compared with DFE-C. Taken together, the fold-based evaluation shows that DFE-C and AFS formed the leading group, with AFS slightly outperforming DFE-C in mean accuracy, while DFE-C offered the most balanced overall performance when both effectiveness and stability are considered.
In addition to the numerical results, the training behavior of the fine-tuned deep learning approaches is illustrated using a graphical representation. The training and validation accuracy and loss curves are presented in Fig. 8, showing the evolution of model performance over the training epochs.
Representative training and validation accuracy and loss curves of the fine-tuned deep learning methods from the 5-fold cross-validation experiments: (a,b) EfficientNetB0 (FTL) and (c,d) EfficientNetB3 (AFS).
To provide a detailed view of the class-wise prediction outcomes, representative confusion matrices were generated for the two selected configurations, namely Approach-2 (DFE-C) and Approach-3 (AFS). The confusion matrices, presented in Fig. 9, illustrate the distribution of correct and incorrect predictions across the considered solar PV defect categories. The visual comparison confirms that both methods achieve strong classification performance across most defect types, with particularly high recognition rates for visually distinctive classes such as Physical-Damage, Snow-Covered, and Electrical-damage. At the same time, the matrices indicate that the remaining misclassifications are concentrated primarily between the Clean and Dusty categories, which exhibit greater visual similarity. These results are consistent with the quantitative findings and further demonstrate that Approach-2 offers a highly stable classification behavior, while Approach-3 achieves similarly strong class-wise predictive performance together with the highest mean accuracy among the evaluated methods.
Normalized confusion matrices of the evaluated approaches under the 5-fold cross-validation framework: (a) DFE-C (EfficientNetB0 features + Linear SVM), and (b) AFS (fine-tuned EfficientNetB3).
This section provides an interpretation of the experimental results presented in Sect. 5 by examining the observed performance patterns in relation to the adopted learning strategies, training methods, and data characteristics. The discussion aims to contextualize the numerical findings without introducing new experimental results, focusing instead on explaining the underlying behaviors of the evaluated approaches.
The overall results indicate that all three evaluated approaches are capable of learning meaningful visual representations for multi-class solar PV defect classification. Despite sharing similar convolutional backbones, the approaches exhibit distinct performance characteristics, suggesting that the learning paradigm and training strategy play a critical role in determining model behavior. Similar observations regarding the impact of transfer learning depth and training method on PV defect detection performance have been reported in recent image-based inspection studies³². This finding suggests that architectural choice alone is insufficient to explain performance differences, particularly in image-based fault detection tasks involving limited and heterogeneous data.
Among the evaluated methods, Approach-2 (DFE-C) and Approach-3 (AFS) achieved closely comparable overall performance, while both clearly outperformed Approach-1 (FTL). Approach-3 attained the highest mean accuracy, whereas Approach-2 yielded the highest macro F1-score by a very small margin and, more importantly, exhibited the lowest standard deviation across folds. These results indicate that, beyond mean predictive performance, model stability should also be considered when assessing the suitability of a method for practical solar PV fault classification tasks. Although all methods were re-evaluated under a common five-fold cross-validation framework, formal statistical significance testing was not included; therefore, the reported differences should be interpreted as descriptive performance differences rather than definitive evidence of statistical superiority. A limitation of the present study is that computational efficiency was not quantitatively assessed in terms of inference time, model size, or deployment feasibility; therefore, deployment-oriented conclusions should be interpreted cautiously.
The feature-based learning approach demonstrates a notably stable performance profile. By employing the convolutional network exclusively as a fixed feature extractor and decoupling deep feature generation from the classification stage, this approach avoids the iterative weight updates associated with end-to-end training. As a result, the model exhibits reduced sensitivity to training dynamics and variations across data partitions. The robustness of this behavior is further supported by the comparatively low variance observed across the cross-validation folds. This pattern is also reflected in Fig. 10, where Approach-2 shows the smallest deviation in both accuracy and macro F1-score among the evaluated approaches.
Mean performance values with standard deviation across five-fold cross-validation.
In contrast, the end-to-end fine-tuning–based approaches display a more dynamic learning behavior. Partial fine-tuning enables the models to adapt higher-level features to photovoltaic-specific defect patterns, but it also introduces a higher degree of sensitivity to the optimization process. This behavior is reflected in the training dynamics illustrated in Fig. 8, where early stopping plays a crucial role in controlling overfitting. The observed convergence trends suggest that extending training beyond certain epochs does not necessarily yield additional generalization benefits, emphasizing the importance of carefully regulated training schedules in fine-tuned deep learning models.
A closer examination of class-wise prediction behavior is provided through the representative confusion matrices presented in Fig. 9 for Approach-2 and Approach-3. These visualizations reveal how prediction accuracy varies across defect categories and highlight specific misclassification patterns. In both cases, the remaining errors are concentrated primarily between visually similar classes, particularly Clean and Dusty, whereas more distinctive defect categories are recognized more reliably. Such patterns are often associated with visual similarities between certain surface conditions and defect appearances, as also reported in recent deep learning–based PV defect identification studies using electroluminescence imagery³³. The confusion matrices therefore serve as important diagnostic tools for understanding not only overall classification accuracy but also the limitations of visual discrimination among defect types.
The application of 5-fold cross-validation to all compared approaches further strengthens the reliability of the reported results. Rather than relying on a single train–test split, cross-validation provides insight into how performance fluctuates across different data partitions. The combination of average metric values and corresponding standard deviations offers a more comprehensive assessment of model consistency. This behavior is further illustrated in Fig. 11, which presents the fold-wise variation of accuracy and macro F1-score across the evaluated methods.
This behavior is further illustrated in Fig. 11, where the fold-wise variation of accuracy and macro F1-score highlights the comparative stability of the evaluated methods across five folds.
Fold-wise performance variation of the evaluated methods across five-fold cross-validation: (a) accuracy and (b) macro F1-score.
Recent studies have also shown that hybrid deep learning structures can provide strong diagnostic performance in industrial maintenance and AIoT applications. For example, metaheuristic-optimized BiLSTM and hybrid CNN-GRU frameworks have been reported for fault diagnosis in industrial air separation stations and firefighting pump systems, respectively^34,35. These studies indicate that combining sequential deep learning models with optimization or hybrid feature-learning strategies can improve diagnostic capability under complex operating conditions. However, such approaches are generally developed for sensor-driven time-series data and AIoT monitoring infrastructures, whereas the present study focuses on image-based multi-class PV fault classification. Therefore, these works are valuable as cross-domain evidence supporting the effectiveness of hybrid deep learning, but they are not directly comparable to the present RGB image-based PV diagnostic framework.
Despite the encouraging outcomes, several limitations of the present study should be acknowledged. In preliminary experiments conducted prior to the final evaluation setup, the influence of dataset size on model performance was observed, where training on a smaller image subset resulted in limited learning capacity, while expanding the dataset led to a substantial improvement in classification performance. This observation underscores the strong dependency of deep learning–based PV fault detection models on data availability and diversity.
The dataset size and class distribution may limit exposure to rare defect patterns, potentially affecting model generalization under unseen image conditions. Another limitation of this study is that it relies only on RGB image data. The inclusion of thermal or electrical data may improve diagnostic completeness and can be considered in future work. Additionally, environmental factors such as illumination variability, camera angle differences, and panel aging effects are not explicitly modeled whereas recent PV inspection frameworks that combine visible and infrared imagery emphasize that such acquisition variability can be addressed through multimodal data capture and robust detection pipelines³⁶. Addressing these factors through larger and more diverse datasets, multi-view imagery, or complementary sensing modalities represents a promising direction for future research.
Furthermore, because this research relies on static datasets, the impact of extreme environmental variability and image acquisition changes was not explicitly modeled. Expanding the dataset to include rare structural anomalies and more diverse environmental conditions may further improve the robustness of the framework in future studies.
Another promising direction for future research is the integration of recently proposed variable-generation and dynamic model-selection methods, such as HCVCM, SVCM, their modified variants, and DSM, into PV defect classification frameworks. In this study, such methods could be applied not to raw images, but to the deep features extracted by convolutional backbones. Their use may help produce more informative and less redundant feature sets before the final classification stage. This may improve class separability and reduce confusion between visually similar defect categories. Dynamic selection strategies may also allow the system to choose the most suitable classifier or decision rule for each input sample. In this way, the complementary strengths of different learning approaches could be used more effectively. Although these methods were not examined in the present study, they represent a valuable direction for future work and may further improve prediction accuracy and generalization performance.
This study has provided a comprehensive evaluation of various deep learning strategies for the automated diagnostic inspection of PV panels. By comparing feature-based learning with end-to-end fine-tuning approaches, the analysis suggests that the effectiveness of a diagnostic framework depends not only on architectural complexity but also on how pre-trained representations are adapted to the target domain. The results indicate that the DFE-C approach achieves a superior overall accuracy of 94.05%, demonstrating that extracting features via EfficientNetB0 combined with a linear classifier provides a stable and reliable solution for photovoltaic fault detection.
The findings indicate that while fine-tuning enhances adaptability to photovoltaic-specific defect characteristics, feature-based learning offers a more stable alternative under limited data conditions. The observed trade-offs between frozen feature extraction and partial parameter updating are aligned with recent findings questioning the universal necessity of extensive pre-training and highlighting the role of domain similarity in transfer performance³⁷. Furthermore, the dependency of model performance on dataset scale and diversity remains a critical factor in achieving robust generalization across different environmental settings.
In conclusion, the transition from experimental AI models to more reliable maintenance-oriented diagnostic frameworks require a balanced focus on both classification accuracy and broader practical considerations. Future research should prioritize the integration of multi-view or multimodal imagery, together with dedicated investigation of lightweight architectures for resource-constrained inspection platforms¹. Addressing these challenges may further support the development of more reliable and scalable AI-based inspection frameworks for solar energy applications.
In light of our experimental observations, the primary focus of future work will involve expanding the current dataset to include underrepresented defect categories that were identified as challenging during the training phase. Since our results indicated that certain visual similarities between texture-based contamination and physical damage lead to occasional misclassifications, we intend to refine the feature extraction process to capture more granular textural details. Furthermore, the variability observed across the cross-validation folds suggests a need for more robust data augmentation techniques that can better represent variations in lighting conditions and camera angles.
Beyond improving standalone classification accuracy, future research should prioritize the integration of these deep learning models into automated aerial inspection frameworks. The transition from offline diagnostic tools to real-time monitoring systems requires the development of lightweight architectures that can maintain high precision on resource-constrained edge devices, such as drones. Future efforts will therefore focus on model compression and quantization techniques to ensure operational feasibility in large-scale PV plants³⁸. Additionally, incorporating multi-spectral or temporal data could provide a more comprehensive diagnostic foundation to effectively distinguish between transient surface anomalies and permanent structural defects.
The datasets used and/or analysed during the current study are available upon request.
Generative adversarial network
Operation and maintenance
Artificial intelligence
Convolutional neural network
Vision transformer
Red, green, blue
Support vector machine
Artificial neural network
High correlated variable creator machine
Stronger variable creator machine
Dynamic selecting machine
Harmonic mean of precision and recall
Input image
Feature extraction function
Model parameters
Deep feature embedding
Number of classes
Loss function
Chao, X., Zhang, L., Li, Y., Huang, C. & Li, J. Photovoltaic fault detection based on infrared and visible image augmentation and fusion. Turk. J. Agric. For. 48(3), 430–442. https://doi.org/10.55730/1300-011X.3192 (2024).
Article  Google Scholar 
Guo, Y., Wang, X. & Lin, Z. RFE-YOLO: A study on photovoltaic module fault detection algorithm based on multimodal feature fusion. Sensors 25(21), 6774. https://doi.org/10.3390/s25216774 (2025).
Article  ADS  PubMed  PubMed Central  Google Scholar 
Ayunts, H., Agaian, S. S. & Grigoryan, A. M. Solar photovoltaic system fault classification via hierarchical deep learning with imbalanced multi-class thermal dataset. Energies 19(2), 462. https://doi.org/10.3390/en19020462 (2026).
Article  Google Scholar 
Lin, P. et al. Fault diagnosis of photovoltaic arrays with different degradation levels based on cross-domain adaptive generative adversarial network. Appl. Energy 386, 122415. https://doi.org/10.1016/j.apenergy.2025.125578 (2025).
Article  Google Scholar 
Ayturan, K., Sarıkamış, B., Akşahin, M. F. & Kutbay, U. SPHERE: Benchmarking YOLO vs. CNN on a novel dataset for high-accuracy solar panel defect detection in renewable energy systems. Appl. Sci. 15(9), 4880. https://doi.org/10.3390/app15094880 (2025).
Article  CAS  Google Scholar 
Duranay, Z. & Demir, K. A deep learning approach for fault detection in photovoltaic systems using MobileNetV3. Gazi Univ. J. Sci. Part. A: Eng. Innov. 12(1), 197–212. https://doi.org/10.54287/gujsa.1596110 (2025).
Article  Google Scholar 
Hamdi, A., Noura, H. N. & Azar, J. Deep learning-based approach to automated monitoring of defects and soiling on solar panels. Future Internet 17(10), 433. https://doi.org/10.3390/fi17100433 (2025).
Article  Google Scholar 
Zwirtes, J., Libano, F. B., Lima, M. R., Silva, L. A. L. & Pignaton de Freitas, E. Fault detection in photovoltaic systems using a machine learning approach. IEEE Access 13, 41406–41421. https://doi.org/10.1109/ACCESS.2025.3547838 (2025).
Article  Google Scholar 
Hussain, M. & Khanam, R. In-depth review of YOLOv1 to YOLOv10 variants for enhanced photovoltaic defect detection. Solar 4(3), 351–386. https://doi.org/10.3390/solar4030016 (2024).
Article  Google Scholar 
Matusz-Kalász, D., Bodnár, I. & Jobbágy, M. An overview of CNN-based image analysis in solar cells, photovoltaic modules, and power plants. Appl. Sci. 15(10), 5511. https://doi.org/10.3390/app15105511 (2025).
Article  CAS  Google Scholar 
Hussain, M., Khanam, R. & Hill, R. Solar panel surface defect and dust detection: Deep learning approach. J. Imaging 11(9), 287. https://doi.org/10.3390/jimaging11090287 (2025).
Article  Google Scholar 
Masita, K., Hasan, A., Shongwe, T. & Abu Hilal, H. Deep learning in defects detection of PV modules: A review. Sol. Energy Adv. 5, 100090. https://doi.org/10.1016/j.seja.2025.100090 (2025).
Article  Google Scholar 
Jimenez, K., Cano, J. B. & Velilla, E. A lightweight deep learning model for fault detection of PV modules using thermal images. Sol. Energy 303, 114120. https://doi.org/10.1016/j.solener.2025.114120 (2026).
Article  Google Scholar 
Setiawan, E. A. & Fathurrahman, M. Exploring insights on deep learning-based photovoltaic fault detection for monofacial and bifacial modules using thermography. International Journal of Cognitive Computing in Engineering 6, 495–507. https://doi.org/10.1016/j.ijcce.2025.04.001 (2025).
Article  Google Scholar 
Sabati, A., Bayindir, R. & Rahebi, J. Photovoltaic panels fault detection with convolutional neural network and bitterling fish optimization (BFO) algorithm. Int. J. Comput. Intell. Syst. 18(1), 239. https://doi.org/10.1007/s44196-025-00984-4 (2025).
Article  Google Scholar 
Duranay, Z. B. Fault detection in solar energy systems: A deep learning approach. Electronics 12(21), 4397. https://doi.org/10.3390/electronics12214397 (2023).
Article  Google Scholar 
Haidari, P., Hajiahmad, A., Jafari, A. & Nasiri, A. Deep learning-based model for fault classification in solar modules using infrared images. Sustain. Energy Technol. Assess. 52, 102110. https://doi.org/10.1016/j.seta.2022.102110 (2022).
Article  Google Scholar 
Ramirez, D. F., Pujara, D., Tepedelenlioglu, C., Srinivasan, D. & Spanias, A. Infrared computer vision for utility-scale photovoltaic array inspection. In 15th International Conference on Information, Intelligence, Systems and Applications (IISA 2024), 1–4. https://doi.org/10.1109/IISA62523.2024.10786613 (IEEE, 2024).
Korkmaz, D. & Acikgoz, H. An efficient fault classification method in solar photovoltaic modules using transfer learning and multi-scale convolutional neural network. Eng. Appl. Artif. Intell. 113, 104959. https://doi.org/10.1016/j.engappai.2022.104959 (2022).
Article  Google Scholar 
Liu, Q., Liu, M., Wang, C. & Wu, Q. M. J. An efficient CNN-based detector for photovoltaic module cells defect detection in electroluminescence images. Sol. Energy 267, 112245. https://doi.org/10.1016/j.solener.2023.112245 (2024).
Article  Google Scholar 
Venkatesh S., N., Sripada, D., Sugumaran, V. & Aghaei, M. Detection of visual faults in photovoltaic modules using a stacking ensemble approach. Heliyon 10(6), e27894. https://doi.org/10.1016/j.heliyon.2024.e27894 (2024).
Article  Google Scholar 
Deitsch, S. et al. Automatic classification of defective photovoltaic module cells in electroluminescence images. Sol. Energy 185, 455–468. https://doi.org/10.1016/j.solener.2019.02.067 (2019).
Article  ADS  Google Scholar 
Rudro, R. A. M. et al. SPF-Net: Solar panel fault detection using U-Net based deep learning image classification. Energy Rep. 12, 1580–1594. https://doi.org/10.1016/j.egyr.2024.07.044 (2024).
Article  Google Scholar 
Johnson, J. M. & Khoshgoftaar, T. M. Survey on deep learning with class imbalance. J. Big Data 6(1), 27. https://doi.org/10.1186/s40537-019-0192-5 (2019).
Article  Google Scholar 
Tan, M. & Le, Q. V. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning (ICML 2019). https://doi.org/10.48550/arXiv.1905.11946 (2019).
Zhang, Y. & Yang, Q. A survey on multi-task learning. IEEE Trans. Knowl. Data Eng. 34(12), 5586–5609. https://doi.org/10.1109/TKDE.2021.3070203 (2021).
Article  Google Scholar 
Shorten, C. & Khoshgoftaar, T. M. A survey on image data augmentation for deep learning. J. Big Data 6(1), 60. https://doi.org/10.1186/s40537-019-0197-0 (2019).
Article  Google Scholar 
Kolesnikov, A. et al. Big Transfer (BiT): General visual representation learning. In European Conference on Computer Vision (ECCV 2020). https://doi.org/10.48550/arXiv.1912.11370 (2020).
Zhuang, F. et al. A comprehensive survey on transfer learning. Proc. IEEE 109(1), 43–76. https://doi.org/10.48550/arXiv.1911.02685 (2021).
Article  ADS  Google Scholar 
Wightman, R., Touvron, H. & Jégou, H. ResNet strikes back: An improved training procedure in timm. IEEE Trans. Pattern Anal. Mach. Intell. https://doi.org/10.48550/arXiv.2110.00476 (2024).
Article  Google Scholar 
Agarwala, A., Pennington, J., Dauphin, Y. & Schoenholz, S. S. Temperature check: theory and practice for training models with softmax-cross-entropy losses. arXiv preprint arXiv:2010.07344. https://doi.org/10.48550/arXiv.2010.07344 (2020).
Akram, M. W., Li, G., Jin, Y., Chen, X. & Zhu, C. CNN based automatic detection of photovoltaic cell defects in electroluminescence images. Energy 189, 116319. https://doi.org/10.1016/j.energy.2019.116319 (2020).
Article  Google Scholar 
Tang, W., Yang, Q., Xiong, K. & Yan, W. Deep learning based automatic defect identification of photovoltaic module using electroluminescence images. Solar Energy 201, 453–460. https://doi.org/10.1016/j.solener.2020.03.049 (2020).
Article  ADS  Google Scholar 
Nguyen, T. P., Lin, C.-C. & Chen, M.-Y. Fault detection of industrial air separation stations based on metaheuristic optimization and bidirectional long short-term memory technique. Internet Things 36, 101850. https://doi.org/10.1016/j.iot.2025.101850 (2025).
Article  Google Scholar 
Nguyen, D.-T., Nguyen, T.-P. & Cho, M.-Y. Cloud-based AIoT intelligent infrastructure for firefighting pump fault diagnosis-based hybrid CNN-GRU deep learning technique. J. Supercomput. 81, Article 462. https://doi.org/10.1007/s11227-025-06965-w (2025).
Article  Google Scholar 
Hong, F. et al. A novel framework on intelligent detection for module defects of PV plant combining the visible and infrared images. Solar Energy 236, 406–416. https://doi.org/10.1016/j.solener.2022.03.018 (2022).
Article  ADS  Google Scholar 
He, K., Girshick, R. & Dollár, P. Rethinking ImageNet pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV 2019), 4918–4927. https://doi.org/10.48550/arXiv.1811.08883 (2019).
Ledmaoui, Y., El Maghraoui, A., El Aroussi, M. & Saadane, R. Enhanced fault detection in photovoltaic panels using CNN-based classificationwith PyQt5 implementation. Sensors 24(22), 7407. https://doi.org/10.3390/s24227407 (2024).
Article  ADS  PubMed  PubMed Central  Google Scholar 
Download references
The datasets used in this study were obtained from publicly available online sources, which provided essential data for performance analysis and model validation. The authors also appreciate the valuable feedback received from colleagues and reviewers.
Energy Systems Engineering, Ankara Yıldırım Beyazıt University, 06090, Ankara, Turkey
Okan Şen, Obed N. Onsomu & Bülent Yeşilata
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Search author on:PubMed Google Scholar
Author contribution is as follows: Conceptualization, O.Ş. and O.N.; methodology, O.Ş and O.N.; software, O.Ş.; validation, O.Ş., O.N. and B.Y.; formal analysis, O.Ş.; investigation, O.N.; resources, B.Y; data curation, O.Ş.; original draft preparation, O.Ş, and O.N.; review and editing, O.N; visualization, O.Ş.; supervision, B.Y.; project administration, B.Y;. All authors have read and agreed to the published version of the manuscript.
Correspondence to Obed N. Onsomu.
The authors declare no competing interests.
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.
Reprints and permissions
Şen, O., Onsomu, O.N. & Yeşilata, B. AI-based visual analysis of photovoltaic panels for fault detection and maintenance support. Sci Rep 16, 21371 (2026). https://doi.org/10.1038/s41598-026-51711-8
Download citation
Received: 23 February 2026
Accepted: 29 April 2026
Published: 10 May 2026
Version of record: 09 July 2026
DOI: https://doi.org/10.1038/s41598-026-51711-8
Anyone you share the following link with will be able to read this content:
Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative
Advertisement
Scientific Reports (Sci Rep)
ISSN 2045-2322 (online)
© 2026 Springer Nature Limited
Sign up for the Nature Briefing: AI and Robotics newsletter — what matters in AI and robotics research, free to your inbox weekly.

source