1. Bouira University Digital Space
2. Thèses Doctorat
3. Faculté des Sciences et des Sciences Appliquées

Title:	Fault diagnosis in a photovoltaic system
Authors:	ACHIT, Mohamed
Keywords:	Photovoltaic (PV) inspection ; Deep learning ; Object detection ; Thermal infrared imaging ; Electroluminescence (EL) ; RT-DETR ; YOLOv8 ; PV-YOLOv12n ; Hotspot detection ; Defect screening
Issue Date:	2026
Publisher:	university of bouira
Citation:	This thesis examined the use of deep learning–based object detection methods to support automated inspection of photovoltaic panels and solar cells using two complementary imaging techniques: thermal infrared and electroluminescence. The overarching goal was to evaluate modern detection architectures, determine their strengths and limitations in different inspection contexts, and propose model improvements where necessary. The detailed experiments conducted throughout the thesis allow several important conclusions to be drawn. First, the study demonstrated that thermal hotspot detection can be effectively performed using lightweight convolutional object detectors. Across all experiments, YOLOv8 consistently outperformed RT-DETR, achieving higher accuracy, faster convergence, and smoother loss behavior. These findings indicate that, for thermal imagery characterized by limited texture and low visual complexity, convolution-based detectors remain highly competitive and often more suitable than transformer-based architectures. Furthermore, the superiority of YOLOv8-l highlights the importance of selecting an architecture with sufficient feature extraction depth to handle variations in panel surface temperature and hotspot shape. Second, the analysis of EL imagery showed that detecting cell-level defects requires models capable of capturing much finer features and subtle intensity variations. The proposed PV YOLOv12n model, developed by introducing a refined feature aggregation block, proved more effective than the baseline YOLOv12n in identifying low-contrast cracks and small structural anomalies. This improvement illustrates that even targeted architectural modifications can yield meaningful performance gains when they align with the characteristics of the imaging modality. The results also suggest that specialized EL-focused detectors can provide valuable contributions to large-scale defect screening. Overall, the research presented in this thesis demonstrates that deep learning offers robust solutions for both panel-level and cell-level PV inspection. The systematic comparison of models, the detailed examination of training behavior, and the introduction of an improved network architecture collectively contribute to a deeper understanding of how current computer vision tools can be adapted for renewable energy applications. In practical terms, the findings support the deployment of real-time inspection systems that can reduce maintenance costs, limit energy losses, and improve the long-term performance of photovoltaic installations.
URI:	http://dspace.univ-bouira.dz:8080/jspui/handle/123456789/20159
Appears in Collections:	Faculté des Sciences et des Sciences Appliquées

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.