Aiming　at　the　problems　of　poor　interpretability　of　traditional　neural　network　and　bloated　traditional　model　in　photovoltaic　panel　defect　detection.　Firstly,　an　FN-v8　model　was　proposed　for　detecting　various　types　of　defects　in　photovoltaic　panels.　The　FN-v8　model　reduces　redundant　computing　and　memory　access,　while　also　reducing　the　GFLOPs　of　the　model.　Then,　Grad-CAM　was　used　to　further　demonstrate　the　stronger　interpretability　of　the　new　model　through　passive　interpretation.　Compared　with　other　traditional　models　with　relatively　bulky　structures,　the　proposed　new　model　has　smaller　parameter　quantities　and　stronger　interpretability.　Experiments　were　conducted　on　devices　with　the　same　performance　and　the　same　dataset　of　photovoltaic　panel　defects.　The　results　showed　that　the　proposed　new　model　has　a　smaller　volume,　faster　calculation　speed,　and　higher　accuracy　than　the　original　model.　The　GFLOPs　of　the　new　model　have　decreased　by　18.25%　compared　to　the　original　model,　significantly　reducing　the　computational　complexity,　and　the　visualization　of　the　thermodynamic　diagram　is　more　accurate,　further　indicating　an　improvement　in　interpretability.　Moreover,　the　new　model　has　better　detection　accuracy　for　photovoltaic　panel　detection,　with　an　average　detection　accuracy(　mAP@0.5　)Improved　by　2.1%　compared　to　the　original　model.　©　2023　SPIE.