Gespeichert in:
| 1. Verfasser: | |
|---|---|
| Format: | Recurso digital |
| Sprache: | |
| Veröffentlicht: |
Zenodo
2026
|
| Schlagworte: | |
| Online-Zugang: | https://doi.org/10.5281/zenodo.20067683 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
Inhaltsangabe:
- <p> </p> <p class="MsoNormal"><strong><em><span>Automated defect detection in photovoltaic (PV) modules is a critical component of renewable energy infrastructure maintenance. However, the efficacy of such systems is frequently limited by the quality of training data rather than architectural constraints. A recurring issue in public repositories is data fragmentation and inconsistent labelling, which impedes the development of generalized models. This study introduces a data-centric methodology focused on the curation and unification of a comprehensive PV thermal dataset from multiple sources. To resolve taxonomic discrepancies such as conflicting definitions of” hotspots” and” cracks”, a binary classification protocol (Defect vs. Background) is adopted. This approach eliminates label noise, establishing a consistent foundation for pre-training. The YOLOv9 architecture and its variants (Gelan-c, Gelan-e, Yolov9-c, and Yolov9-e) are employed to validate the dataset’s integrity. Empirical analysis confirms that this unified strategy effectively captures fundamental defect characteristics, providing a robust baseline for future PV monitoring applications.</span></em></strong></p>