Multimodal Transformer Defect Recognition Dataset

Overview This dataset supports multimodal defect recognition for electrical (main) transformers. It comprises three heterogeneous modalities namely, inspection-log text, dissolved gas analysis (DGA) measurements, and appearance images, aligned to a unified ten-class defect recognition framework. The unified defect recognition framework is: Θ = {overheating, discharge, moisture ingress, solid-insulation aging, fire, oil leakage, rust and corrosion, cable defect, insulator crack, normal} Each modality covers only a partially overlapping subset of these ten classes; this partial coverage is the phenomenon the accompanying paper's capability-mask mechanism is designed to handle.