Identifying　fault　sections　in　single-phase　ground　(SPG)　faults　is　essential　for　electric　utilities　to　promptly　isolate　faults　and　restore　service.　A　deep　learning-based　approach　leveraging　feature　fusion　has　been　proposed　for　SPG　fault　section　location,　utilizing　transient　zero-sequence　currents　(TZSCs)　captured　by　feeder　terminal　units　(FTUs).　Initially,　a　convolutional　neural　network　(ConvNet)　is　pre-trained　on　TZSC　waveforms　to　distinguish　data　from　the　upstream　and　downstream　of　the　fault　point,　acting　as　a　feature　extractor.　This　pre-training　enables　the　model　to　capture　distinct　transient　characteristics　from　both　ends　of　the　fault.　The　pre-trained　ConvNet　is　then　replicated　to　form　a　dual-branch　architecture,　where　TZSC　data　from　both　ends　of　the　feeder　section　are　input　into　the　respective　branches.　The　features　extracted　from　these　branches　are　concatenated　at　a　fusion　layer,　allowing　the　model　to　effectively　integrate　the　transient　information　from　upstream　and　downstream,　leading　to　more　precise　fault　section　location.　Compared　with　existing　methods,　our　approach　demonstrates　robustness　under　various　conditions,　including　simulation　verification　and　field　verification.　Extensive　testing　shows　that　the　model　maintains　high　performance　even　with　limited　field　data,　and　fine-tuning　further　enhances　its　practical　applicability　for　engineering.　Moreover,　an　industrial　prototype　utilizing　Raspberry　Pi　4B　has　been　implemented　in　real-world　distribution　networks,　where　fault　data　are　transmitted　to　the　main　station,　further　optimizing　the　fault　section　location　process　using　our　proposed　approach.