Bearing　fault　diagnosis　is　an　important　research　field　for　rotating　machinery　health　monitoring.　Recently,　many　intelligent　fault　diagnosis　methods　driven　by　big　data,　such　as　transfer　learning,　have　been　studied.　However,　there　are　two　shortcomings　for　the　prior　transfer　learning　method　in　industry　application.　First,　it　is　necessary　to　design　a　complex　loss　function　to　enhance　the　similarity　between　the　two　domains　further.　Second,　previous　studies　required　big　data　both　in　source　and　target　task,　without　considering　the　lack　of　sufficient　training　samples.　Inspired　by　relevant　research　work,　this　article　proposes　a　local　joint　distribution　discrepancy　to　increase　similar　features.　A　sub-domain　adaptive　transfer　learning　is　designed　to　detect　bearing　faults　based　on　the　residual　network.　Two　kinds　of　transfer　experiments　are　designed　to　verify　the　method　effectiveness.　After　that,　the　impact　of　small　training　samples　and　noise　on　the　results　is　explored.　The　proposed　method　reaches　high　accuracy.