Porcelain Disk type insulators (Classification and Detection task)
Description
This dataset contains 255 aerial images of porcelain disc insulators photographed on energized 500 kV overhead transmission lines in Sindh Province, Pakistan, together with their corresponding instance-segmentation annotations. Images were collected in 2024–2025 along four transmission circuits — NKI–Jamshoro (single circuit), Dadu–Jamshoro (single circuit), CPHGC–Jamshoro (double circuit), and K2 K3–Jamshoro Circuit-1 (double circuit) — using a DJI Air 2S quadcopter drone (1-inch CMOS sensor, 20-megapixel stills, up to 5.4K video). Photographs were taken directly during flight and, in some cases, extracted as still frames from raw video recordings. A minimum stand-off distance of approximately 1 metre from insulators and live conductors was maintained throughout to avoid electrical discharge to the aircraft. Each image was manually annotated using the Roboflow platform with polygon (instance-segmentation) masks, labelling every visible insulator unit or string into one of four condition classes: Insulator (healthy, no visible defect), Broken (cracked, chipped, or shattered porcelain), Dirty/Flashed (surface contamination or visible flashover/arcing damage), and Shaded (partially obscured or silhouetted by shadow/backlighting). In total, the dataset contains 973 annotated instances (mean 3.82 annotations per image), split into training (768 annotations), validation (108 annotations), and test (97 annotations) subsets. The data are provided in the Ultralytics YOLOv8 instance-segmentation export format: per-image polygon-label text files, an accompanying data.yaml configuration file (class order: Broken, Dirty/Flashed, Insulator, Shaded), and README files documenting the export source. The dataset is intended to support training and benchmarking of deep-learning object-detection and instance-segmentation models for automated, drone-based visual inspection and condition monitoring of high-voltage transmission-line insulators. Released under a CC BY 4.0 licence.
Files
Institutions
- Mehran University of Engineering and TechnologySindh, Jamshoro