A machine learning framework that combines patch-based image classification, synthetic data generation, and statistical feature engineering has achieved an area under the precision-recall curve of 0.9526 in identifying rare bright spot defects from electroluminescence images of photovoltaic solar panels. Developed to address critical reliability challenges during panel fabrication, the approach transforms complex image-wide defect searches into localized patch evaluations, demonstrating high precision and recall on real-world manufacturing data. Early detection of structural irregularities remains vital to ensure that photovoltaic modules operate safely and maintain anticipated energy yields across their operational lifespans.
Bright spot anomalies typically originate during the final lamination stage of photovoltaic panel manufacturing, where individual module layers are evacuated of air, heated, and mechanically pressed together to secure solar cells in place. Under the mechanical pressure and elevated temperatures of this lamination procedure, surface electrical wiring can make unintended contact, creating localized pathways of unusually low electrical resistance. When panels undergo electroluminescence quality inspections, these low-resistance zones appear as intense white regions, predominantly located around the corners of individual photovoltaic cells. If left uncorrected prior to deployment, bright spots cause short circuits in operational arrays, generating excessive localized energy, extreme overheating, and severe fire hazards.
Detecting bright spots presents substantial technical difficulties because they occur infrequently during normal production runs, leading to severe class imbalance in automated inspection datasets. An evaluation of 225 electroluminescence panel images obtained from an active, operational solar panel manufacturing facility revealed that only 32 panels—amounting to approximately 14 percent—contained confirmed bright spots, while 193 panels exhibited no defect. Because missing a bright spot poses severe safety risks to life and property while threatening manufacturer credibility, automated computer vision systems must reliably detect anomalies without generating false negatives, a task where standard deep learning models frequently fail when confronted with sparse positive instances.
To establish baseline benchmarks, researchers first tested standard object detection architectures, including YOLOv8, Faster R-CNN, Single Shot MultiBox Detector (SSD), and RetinaNet, directly on full photovoltaic panel images annotated by subject matter experts. Operating with resized dimensions of 1600 pixels, a batch size of 8, and a training span of 100 epochs, none of the full-panel models reliably detected bright spots in stratified test sets. While Faster R-CNN attained a higher precision-recall curve area than competing baselines, it registered a very poor intersection-over-union score, revealing an unacceptably high rate of false positive detections across the inspection field.
Recognizing the limitations of full-panel detection, the researchers partitioned the panel images into isolated photovoltaic cells, reducing the image input size to 320 pixels while retaining the original training hyperparameters. Although individual cells represent smaller, more uniform visual areas than complete modules, retraining the object detection architectures at the cell level still failed to yield performance metrics adequate for production deployment. The failure of bounding-box localization models at both the panel and cell levels prompted the team to restructure the computational formulation of the inspection task entirely.
The study addressed these limitations by converting the detection problem into a binary classification task, segmenting each extracted photovoltaic cell into small, non-overlapping visual patches whose dimensions mirrored standard annotation bounding boxes. Because the precise cell row, cell column, and spatial coordinates of each patch within the module are systematically recorded, classifying an individual patch as positive or negative simultaneously detects and localizes defects across the broader panel. However, dividing cells into thousands of discrete patches drastically intensified the underlying class imbalance, rendering basic undersampling strategies ineffective and causing conventional oversampling methods to falter.
To resolve the severe patch-level data imbalance, the researchers deployed generative adversarial networks (GANs) to generate 2,000 synthetic electroluminescence patches containing bright spots modeled on authentic defect samples. Unlike manual oversampling or synthetic minority over-sampling technique (SMOTE) algorithms—both of which failed to enhance detection performance in comparative trials—the generative network learned the underlying statistical patterns of genuine bright spots to produce realistic synthetic training samples. Incorporating these synthetic instances balanced the training distribution, mitigated classification bias toward defect-free patches, and provided subsequent models with sufficient positive examples to discern subtle anomalies.
In parallel with synthetic data generation, the investigation extracted novel discriminative features derived from the statistical distribution of pixel intensities within each patch. In standard eight-bit electroluminescence imagery, pixel brightness ranges from zero for completely dark areas to 255 for white regions, with bright spot defects appearing as statistical outliers concentrated near peak brightness. Analyzing the gradient of the Inverse Cumulative Density Function (ICDF) across patch percentiles revealed that patches containing bright spots exhibited steep slope increases after the 95th percentile, displaying an average ICDF slope of 6.92 compared to just 0.68 in defect-free patches. These gradient values, along with maximum intensity readings and threshold flags, formed a compact feature set optimized via Bayesian hyperparameter tuning.
When evaluated across diverse algorithmic frameworks, traditional classifiers trained on these novel features and GAN-augmented data significantly outperformed advanced deep learning baselines. Unaugmented deep architectures and representation models—including VGG19, ResNet-50, GoogleNet, vanilla convolutional networks, and autoencoders—all failed to achieve satisfactory precision-recall metrics. In contrast, an XGBoost classifier leveraging the novel statistical features achieved an area under the precision-recall curve of 0.8871 on baseline patch data, which subsequently rose to a peak score of 0.9502 when combined with GAN oversampling, confirming the diagnostic value of targeted statistical indicators over raw pixel representations.
While the combination of generative oversampling and patch-level gradient analysis demonstrated high efficacy, the authors noted several boundaries and avenues for refining the methodology. Anomaly detection models such as convolutional and variational autoencoders trained exclusively on regular, defect-free patches proved incapable of reliably flagging bright spot outliers. Future work aims to optimize generative data pipelines by assessing synthetic image quality through average log-likelihood and inception scores, tuning the volume of synthetic samples, and validating the patch classification framework across diverse commercial photovoltaic panel architectures to ensure robust operational generalization.
Beyond localized electrical bridging, quality control during photovoltaic fabrication must continuously guard against structural flaws such as micro-cracks, material voids, and chemical impurities. These physical imperfections degrade operational durability, reduce aggregate electrical generation, and shorten panel lifespan as solar deployment expands to replace greenhouse-gas-emitting fossil fuels. At the structural level, an individual photovoltaic cell serves as the fundamental building block of a solar module, converting sunlight into electricity. In commercial panels, multiple cells are wired together in series or parallel configurations to accumulate sufficient voltage and current to satisfy electrical grid standards or specialized consumption requirements. Because each module integrates dozens of interconnected cells, defect prevention at the individual cell and patch level remains indispensable for sustaining uniform electrical throughput.
The research also addresses gaps within earlier machine vision literature concerning automated photovoltaic inspection. Prior studies examined diverse computer vision methods on electroluminescence datasets, deploying pipelines such as KAZE descriptors matched with vector of locally aggregated descriptors encoding and support vector machines, alongside deep architectures like ResNet-50 and VGG19. Other researchers introduced bi-directional attention feature pyramid networks or near-infrared Faster R-CNN frameworks to distinguish small defects against intricate backgrounds, while separate teams applied U-Net segmentation to isolate bright spots caused specifically by surface shielding. Similarly, YOLO variants were applied to surface defects, and GoogleNet models were coupled with thermal imaging to detect operational hot spots. However, thermal inspections rely on temperature differentials rather than electroluminescence signals, and none of these published frameworks directly tackled the specific challenge of identifying manufacturing bright spots in panel electroluminescence imagery.
Building upon a preliminary methodology, the authors expanded their benchmarking protocol to incorporate autoencoders alongside traditional and modern oversampling mechanisms. Crucially, the investigation transitioned its primary evaluation metric from the traditional F1 score to the area under the precision-recall curve. While both analytical metrics rely directly on precision and recall relationships, the precision-recall curve provides a substantially more informative evaluation when assessing models over heavily skewed distributions. By evaluating classification behavior across the full spectrum of operating thresholds rather than at a single cutoff point, the precision-recall curve exposes how effectively an algorithm discriminates the rare minority class.
To implement this framework, the factory dataset of 225 panel images was partitioned into training, validation, and test subsets comprising 50 percent, 10 percent, and 40 percent of the total data, respectively. Stratified sampling preserved the exact 14 percent defect ratio across all three data splits. The training partition enabled model learning, the 10 percent validation set supported hyperparameter tuning through Bayesian optimization using the Hyperopt library, and the 40 percent test holdout provided an unbiased assessment of generalization capability. Hyperopt tuned classifier parameters as well as specific feature configurations for Logistic Regression, Random Forest, and XGBoost classifiers, evaluating the mathematical characteristics of Inverse Cumulative Density Function curves across varying intensity percentiles.
Comparative trials further demonstrated that deep neural architectures could not overcome data scarcity through synthetic oversampling alone. When convolutional networks and feature representations, including ORB and KAZE with VLAD encoding, VGG19, Vanilla CNN, GoogleNet, and ResNet-50, were retrained on GAN-oversampled patches, none achieved notable performance gains. This outcome underscores the diagnostic superiority of domain-tailored statistical features over raw deep representations. Finally, data governance requirements govern access to the underlying experimental material. Due to intellectual property protections, the 225 industrial electroluminescence images are not publicly accessible and require formal institutional clearance from Reliance Industries Ltd. before being released upon reasonable request by the corresponding author.
Subject of Research: Automated detection of bright spot defects in photovoltaic panel electroluminescence images
Article Title: Accurate detection of bright spots in electro-luminescence images of photovoltaic panels
Article References: Nair, R., Thirwani, A., & Kulkarni, K. (2025). Accurate detection of bright spots in electro-luminescence images of photovoltaic panels. Discover Data, 3(1), Article 42. https://doi.org/10.1007/s44248-025-00085-9
Image Credits: AI Generated
DOI: 10.1007/s44248-025-00085-9
Keywords: photovoltaic panels, electroluminescence imaging, bright spot detection, solar cell defects, generative adversarial networks, patch extraction, pixel intensity distribution, inverse cumulative density function, XGBoost classifier, precision-recall curve, class imbalance, manufacturing quality control, object detection benchmarking, synthetic data augmentation, binary image classification, Bayesian hyperparameter optimization, solar energy reliability
News Source: Denise Maddox. (November 18, 2026). Patch Classification and Generative Models Accurately Detect Photovoltaic Panel Defects. Scienmag.



