Browsing by Author "Rocha, Pedro Daniel Carvalheiro"
Now showing 1 - 1 of 1
Results Per Page
Sort Options
- Automated visual inspection of electrical grid assets using deep learningPublication . Rocha, Pedro Daniel Carvalheiro; Lopes, Fernando José PimentelThe global growing demand for electric power, projected to double by 2050, requires extensive upgrades in both Power Generation and Transport and Distribution (T&D) systems. To address these needs, new infrastructures are required. Meanwhile, the reliability of the aging electrical generation and distribution systems is critical. On the one hand, one main direction for power generation is the increased use of Photovoltaic (PV) technologies, with the global solar power capacity exceeding 1 TW in 2022. With the rapid expansion of PV systems, manual image collection for asset inspection becomes increasingly impractical, leading to imagery collected by UAVs. On the other hand, in the next three decades, worldwide T&D systems may exceed 160 million lan, where High Voltage (HV) power line insulators account for over 50% of the maintenance costs. Electric utility companies are increasingly collecting visual data as part of their inspection process of electrical T &D infrastructures. In both PV power generation and T&D systems, the increasing volume of collected data and its processing, is currently being limited by human interpretation tasks, mainly because escalating this pipeline element trough training is very expensive. However, most inspection tasks with minor responsibility can be automated using computer vision and deep learning techniques to support the system’s growing demand. This work demonstrates the potential of computer vision and deep learning for automated visual inspection of electrical grid assets, validated with real-world data, to reveal advantages and limitations. The project reviews recent advancements in automated visual inspection of electrical assets, namely PV panels and HV insulators, and outlines the essential knowledge in asset inspection, computer vision and deep learning. It also introduces a set of tools currently used to deploy deep learning projects. The developed work includes example applications on automated visual inspection of PV Cells, PV Modules and HV Insulators. PV Cells’ operational condition is classified by processing EL images using the VGG19, resulting in AUC values of 0.95,0.94, and 0.90 for the poly, mono, and mixed datasets, respectively, with overall improvements compared to those reported in the literature. IR images of PV Modules are analyzed using four custom shallow CNNs with varying levels of complexity to classify thermographic patterns into predefined defective classes, achieving Precision, Recall, and F1-score values of 0.87,0.86, and 0.86, respectively. Defects in HV insulators are detected using two SOTA deep learning models with visible light images. YOLOv8s achieves a mAP@50 of 87.9% while Faster R-CNN X101-FPN achieves 87.2% for the same metric. This work can serve as an encouragement in developing a robust model, allowing utility companies to benefit from higher efficiency inspection processes.