2026-10-13T10:31:55Z https://www.scidb.cn/oai

10.57760/sciencedb.46313 2026-08-17T11:23:17Z

2026-08-17 Multimodal Visual Instruction Dataset for Intelligent Transmission Line Inspection: Multi-Turn Question Answering and Risk Analysis Edition doi:10.57760/sciencedb.46313 en This dataset was developed for intelligent transmission-line inspection, multimodal visual question answering, and risk analysis of power equipment defects. The multi-turn question answering and risk analysis edition contains 9,445 image-conversation samples, including 7,599 training samples and 1,846 validation samples. The training split consists of 5,765 original samples retained after duplicate and data-leakage screening and 1,834 class-aware augmented samples.The dataset covers balloons, bird nests, kites, small transmission towers, transmission towers, insulators, broken insulators, flashed insulators, insulators in good condition, damaged conductor insulation, broken conductor strands, and loose conductor strands. The images include different viewing distances, camera angles, weather conditions, lighting conditions, and surrounding environments, reflecting scale variation, partial occlusion, and complex backgrounds encountered in practical transmission-line inspection.Each sample contains one inspection image and three consecutive rounds of question answering, comprising three user instructions and three corresponding responses. The first round focuses on the visible environment, power equipment, and identifiable objects. The second round examines equipment condition, anomaly location, and appearance characteristics. Based on the preceding visual evidence, the third round evaluates potential electrical or mechanical risks and provides maintenance recommendations, such as on-site verification, defect recording, continued monitoring, scheduled maintenance, or component replacement. Contextual continuity is maintained across the conversation, allowing the model to progress from basic visual understanding to defect assessment and professional inspection reasoning.The dataset follows a conversational instruction format suitable for fine-tuning multimodal large language models such as Qwen-VL. All images are referenced using relative paths to facilitate data migration, model training, and experiment reproduction. Before release, the dataset was checked for image-annotation pairing, file readability, duplicate records, train-validation leakage, category validity, image-text semantic consistency, conversation structure, and path portability. The conversations do not contain object-detection procedures, region coordinates, or bounding-box descriptions. The dataset can support research on multi-turn visual question answering, power equipment defect recognition, condition assessment, risk reasoning, maintenance decision support, multimodal instruction tuning, and intelligent inspection report generation. transmission line inspection; multimodal visual question answering; multi-turn dialogue; visual instruction tuning; power equipment defects; risk analysis; insulator defects; conductor defects Xie Xiaowei Cao Hui Dou Liyang Zhang Jingxuan PUBLIC https://creativecommons.org/licenses/by-nc/4.0/ dataset Science Data Bank