Detecting　defects　in　insulators　is　critical　to　ensuring　transmission　reliability　and　reducing　safety　risks.　Traditional　defect　detection　methods　rely　on　manual　feature　engineering　and　classical　algorithms,　which　often　struggle　to　ensure　robustness　in　complex　environments.　The　advent　of　deep　learning　has　revolutionised　insulator　defect　detection　methods　by　providing　superior　feature　extraction　and　real-time　performance.　However,　existing　deep　learning　models　face　challenges　in　detecting　subtle　defects.　To　address　these　limitations,　this　paper　proposes　a　Self-Attention　Transformer　(SATF)　detection　model　tailored　for　embedded　device　applications　such　as　drones.The　SATF　model　incorporates　a　visual　transformer　in　the　multilevel　CNN　architecture　to　enhance　feature　representation,　a　fast　spatial　pyramid　structure　for　uniform　scale　information　capture,　and　a　deformable　self-attention　transformer　module　to　improve　small　object　feature　understanding　of　small　object　features.　The　experimental　results　show　that　the　SATF　model　can　detect　flashover　and　damaged　insulators　with　higher　accuracy　than　YOLOv8s　by　4.2　%　and　YOLO11s　by　3.8　%　while　maintaining　the　computational　efficiency,　which　provides　a　solution　for　automatic　transmission　line　detection.　©　2025　IEEE.