Source-linked AI summary
Automatic Classification of Defective Photovoltaic Module Cells in Electroluminescence Images
Sergiu Deitsch, Vincent Christlein, Stephan Berger, Claudia Buerhop-Lutz, Andreas Maier, Florian Gallwitz, Christian Riess
TL;DR
Defect identification in PV cells is difficult to perform visually and does not scale through direct electrical measurements. The paper develops SVM and CNN classification pipelines for EL images, with the CNN achieving higher average accuracy while the SVM offers a lower-resource alternative.
Problem
Visual defect identification is difficult because efficiency-reducing defects may be invisible, while direct electrical measurements require manual diagnosis and do not scale well.
Method
The paper trains an SVM using KAZE/VGG features and a fine-tuned VGG-19 regression CNN on augmented EL cell images, incorporating labeler confidence.
Results
The CNN reaches 88.42% average accuracy versus 82.44% for the best SVM pipeline, with the CNN gaining about 6% accuracy on polycrystalline cells.
Takeaways & Limitations
Both classifiers are useful for visual inspection; the CNN is preferred when GPUs and higher processing times are available, while the SVM suits low-resource applications.
Takeaways & Limitations
Each solar cell is examined independently, so accurately classifying larger-scale surface effects requires incorporating context across cells.
Abstract
from arXiv · showhide
Electroluminescence (EL) imaging is a useful modality for the inspection of photovoltaic (PV) modules. EL images provide high spatial resolution, which makes it possible to detect even finest defects on the surface of PV modules. However, the analysis of EL images is typically a manual process that is expensive, time-consuming, and requires expert knowledge of many different types of defects. In this work, we investigate two approaches for automatic detection of such defects in a single image of a PV cell. The approaches differ in their hardware requirements, which are dictated by their respective application scenarios. The more hardware-efficient approach is based on hand-crafted features that are classified in a Support Vector Machine (SVM). To obtain a strong performance, we investigate and compare various processing variants. The more hardware-demanding approach uses an end-to-end deep Convolutional Neural Network (CNN) that runs on a Graphics Processing Unit (GPU). Both approaches are trained on 1,968 cells extracted from high resolution EL intensity images of mono- and polycrystalline PV modules. The CNN is more accurate, and reaches an average accuracy of 88.42%. The SVM achieves a slightly lower average accuracy of 82.44%, but can run on arbitrary hardware. Both automated approaches make continuous, highly accurate monitoring of PV cells feasible.
1. Introduction
PV-module defects can reduce efficiency yet remain difficult to assess visually or through scalable direct measurements. The paper proposes per-cell SVM and CNN pipelines for automated defect-likelihood classification from high-resolution EL images.
- Mechanical damage and manufacturing errors can produce defects that decrease solar-module power efficiency.
- Visual inspection is difficult because efficiency-reducing defects may be invisible, while visible defects may not reduce module efficiency.
- Direct electrical measurements are labor-intensive, do not scale well to large plants, and capture only one point in time.
- EL imaging provides higher spatial resolution than IR imaging, enabling detection of small defects such as microcracks.
- The paper targets per-cell defect likelihood rather than exact defect localization because cell-level quality assessment prioritizes identifying areas that may lose efficiency.
- The proposed approaches classify cells using either hand-crafted EL-image features with an SVM or image pixels with a CNN.
2. Related Work
Prior EL-inspection studies often target particular defect types, while broader defect and power-loss prediction remains less addressed. The paper situates its SVM–CNN comparison within this gap and related vision-based inspection work.
- Earlier EL-image studies commonly detect specific intrinsic or extrinsic defects rather than defects expected to reduce module power efficiency.
- Finger interruptions can be confused with cracks even though they may not significantly affect power loss, complicating defect interpretation.
- Existing methods separately address cracks, finger interruptions, or disconnected cells, limiting coverage across defect types.
- The paper reports no previously proposed CNN architecture specifically for EL images, although CNNs have been applied to other inspection modalities.
- Figure 3 frames the SVM alternative as a pipeline with four preprocessing and feature-extraction variations.
3. Methodology
The methodology analyzes individually segmented, size-normalized PV-cell EL images using either an SVM pipeline with engineered local features or a modified CNN. The SVM pipeline evaluates masking, sampling, feature combinations, VLAD encoding, and staged CNN fine-tuning.
- Cell preprocessing: PV modules are subdivided into individual solar cells, increasing training samples while preserving the smallest meaningful mechanically interconnected unit.Cell images are size-normalized to 300 × 300 pixels before analysis.
- SVM classification: The SVM pipeline extracts local descriptors, encodes them globally, and classifies each cell as defective or functional.The pipeline consists of masking, keypoint detection, feature description, encoding, and classification.
- SVM classification: Masking separates cell interiors from backgrounds containing busbars and inter-cell borders, but evaluation finds only minor performance improvements.The mask strictly limits feature extraction to the cell interior.
- Feature extraction: Feature locations are obtained through dense sampling or keypoint detection, using combinations of SIFT, SURF, KAZE, AGAST, PHOW, HOG, and VGG descriptors.Dense sampling uses a fixed grid, whereas keypoint detectors depend on image texture; SIFT and SURF detectors were unreliable in homogeneous monocrystalline cells.
- Model evaluation: The study evaluates feature-detector and descriptor combinations, omits binary descriptors, and reports that staged CNN fine-tuning slightly improves generalization.Twelve feature combinations are evaluated overall, while binary descriptors are excluded because they generally do not outperform real-valued descriptors.
- Feature encoding: VLAD forms a fixed-length global descriptor by aggregating local-descriptor residuals relative to nearest dictionary anchor points.The resulting representation concatenates residual terms into a Kd-dimensional vector and is followed by normalization.
4. Evaluation
The evaluation compares dense and keypoint-based feature configurations for SVM classification, then benchmarks the strongest SVM variants against a CNN across mono- and polycrystalline cells. The CNN generally performs better, while the SVM offers a hardware-efficient alternative with competitive results in some settings.
- Dataset: The dataset contains 2,624 300 × 300-pixel solar-cell images extracted from 44 monocrystalline and polycrystalline PV modules.The images were acquired under controlled manufacturing conditions to reduce illumination and image-quality variation.
- Dataset: Labels were based on expert visual assessment focused on defects associated with power loss above 3%, with lower weights for non-confident decisions.Direct power-degradation measurements were unavailable; non-confident functional and defective assessments received weights of 33% and 67%, respectively.
- Dense Sampling: The strongest dense-sampling configuration used VGG features with a 65 × 65 grid, a linear SVM, sample weighting, and masking.SIFT performed best at 60 × 60 with a weighted linear SVM without masking, while SURF peaked at 70 × 70 with a weighted RBF SVM without masking.
- Keypoint Features: Keypoint-based KAZE/VGG achieved the highest AUC across all modules at 88.51%, followed by KAZE/SIFT at 87.22%.Overall, keypoint features performed better than dense sampling.
- SVM versus CNN: The CNN outperformed the SVM overall, but their total performance difference was only about 6%.For monocrystalline modules, performance was similar; for polycrystalline modules, the CNN was almost 11% better in AUC, where texture variation made classification more difficult.
- SVM versus CNN: The SVM remains useful for on-the-spot PV-module assessment because it can run on commodity hardware without specialized CNN hardware.The proposed evaluation compares feature descriptors, classifiers, weighting, masking, and sampling strategies before selecting SVM configurations for CNN comparison.
KAZE/VGG CNN
The proposed SVM and CNN pipelines are evaluated across training-set sizes, feature-space structure, qualitative explanations, and computational costs. The CNN generally offers stronger performance and faster inference, while the SVM remains more efficient to train and deploy on limited hardware.
- CNN feature space: The CNN feature space separates definite defects across wafer types, whereas functional-cell clusters depend more strongly on the source module type.Definite defective cells form one elongated cluster, while functional cells form type-dependent clusters with additional appearance-related branches.
- CNN feature space: The possibly defective and likely defective categories overlap substantially because they are smaller and more ambiguous than the high-confidence categories.These categories correspond to 33% and 67% defect likelihoods.
- Qualitative results: CAMs can highlight potentially defective regions even when the CNN incorrectly classifies a defective cell as functional.The visual explanations can support inspection in complicated cases, although finger interruptions are not always clearly distinguished from actual defects.
- Computational cost: CNN training takes about 5 hours versus roughly 30 minutes for the SVM, while CNN inference takes just under 20 seconds versus over 8 minutes for the SVM.The reported inference comparison uses GPU execution for the CNN and CPU execution for the SVM.
- Computational cost: Replacing VGG with SIFT or PHOW reduces SVM feature-extraction time from 8 minutes to about 23 or 12 seconds while maintaining similar classification performance.These alternatives target applications requiring both low resource use and fast execution.
- Classification performance: 88.42% accuracy is achieved by the CNN, compared with 82.44% for the best KAZE/VGG linear-SVM pipeline.The corresponding F1 scores are 88.39% for the CNN and 82.52% for the SVM.
- Limitations: The method examines each solar cell independently, so repetitive surface abnormalities requiring module-scale context remain outside its scope.Context-aware classification is identified as future work.
5. Conclusions
The paper presents SVM and CNN frameworks for identifying defective solar cells in high-resolution EL images. The CNN performs better overall, while the SVM offers a lower-resource alternative, with the preferred choice depending on available hardware and processing constraints.
- Contributions: The proposed frameworks use hand-crafted features with an SVM and a fine-tuned VGG-19 regression CNN, both incorporating label confidence during training.The CNN is trained on augmented cell images, while the best SVM pipeline uses KAZE/VGG features in a linear SVM.
- Evaluation: On monocrystalline modules, both classifiers perform similarly, with only a slight average advantage for the CNN.The larger performance difference appears on polycrystalline cells.
- Practical implications: The CNN is preferred when GPUs and longer processing times are available; otherwise, the SVM is viable for low-resource applications.The conclusion frames the choice as dependent on the application scenario.