High-altitude　Operations　in　power　grids　are　numerous,　and　violations　during　these　activities　can　easily　result　in　serious　accidents,　such　as　falls　from　heights.　This　study　aims　to　develop　a　lightweight,　high-precision　model　for　detecting　violations　in　high-altitude　Operations,　enabling　the　efficient　identification　of　unsale　behaviors.　The　research　begins　by　developing　a　foundational　network　model　based　on　the　YOLOv8　algorithm　framework.　Subsequently,　a　Ghost　module　is　integrated　into　the　neck　network,　substituting　Standard　convolution　with　Ghost　Conv.　Additionally,　the　bottleneck　layer　in　the　C2f　module　is　replaced　with　a　Ghost　Bottleneck　layer,　resulting　in　the　creation　of　the　GhostC2f　module,　which　further　enhances　the　neck　network　to　construct　a　lightweight　model.　Subsequently,　an　Efficient　Multi-Scale　Attention　(EMA)　module　is　integrated　into　the　backbone　network　by　embedding　it　within　the　bottleneck　layer　of　the　C2f　module,　resulting　in　the　proposed　C2f_EMA　module.　This　C2f_EMA　module　replaces　the　original　C21　module　in　the　backbone　network,　thereby　enhancing　the　network＇s　ability　to　extract　key　feature　information.　This　approach　enhances　the　model＇s　accuracy　while　maintaining　a　lightweight　loundation.　Finally,　an　Analytic　Network　Process　(ANP)　model　is　constructed　using　the　Analytic　Network　Process　method　and　the　Fuzzy　Comprehensive　Evaluation　method,　allowing　for　the　calculation　of　the　weights　of　the　evaluation　indicators　within　the　model.　Next,　qualitative　evaluations　are　converted　into　quantitative　assessments　to　measure　the　level　of　risk　in　specific　operational　scenarios.　This　process　ultimately　integrates　isolated　unsafe　behaviors　into　a　comprehensive　evaluation　System　for　high-altitude　Operations.　The　results　indicate　that,　compared　to　the　basic　model,　this　new　model　achieves　a　3.3%　increase　in　accuracy,　a　9.　9%　improvement　in　recall　rate,　and　an　8.　4%　enhancement　in　mean　Average　Precision　at　mAP@　50,　along　with　a　4.　6%　increase　at　mAP@50;95.　In　terms　of　lightweight　Performance,　the　number　of　parameters　is　reduced　by　17.　6%,　computational　demand　is　decreased　by　11.　1%,　and　the　model＇s　storage　requirements　are　lowered　by　16.　1%.　Additionally,　the　incidence　of　missed　detections　has　declined.　©　2025　Science　China　Press.　All　rights　reserved.