In　recent　years,　with　the　in-depth　development　of　intelligent　technology,　the　application　of　deep　learning　combined　with　domain　adaptive　methods　in　mechanical　cross-domain　fault　diagnosis　has　increased.　However,　in　the　actual　environment,　the　state　categories　of　the　target　domain　that　need　to　be　tested　may　only　constitute　a　subset　of　the　source　domain.　A　partial　fault　diagnosis　method　for　rolling　bearings　is　proposed　to　aim　at　the　partial　domain　adaptive　problem　in　mechanical　fault　classification.　The　method　introduces　the　temperature　decay　strategy　into　the　traditional　softmax　function,　sharpens　the　weight　distribution　difference　between　the　shared　and　outlier　classes,　uses　the　obtained　class-level　weights　to　weigh　the　source　samples,　and　combines　with　the　confrontation　strategy　to　achieve　domain　adaption.　In　addition,　a　correction　function　is　added　to　the　fault　classifier　to　improve　the　classification　accuracy　of　the　source　domain　samples.　Three　different　bearing　data　sets　were　used　to　verify　the　effect　of　the　proposed　method.　Finally,　the　interpretability　of　the　proposed　model　is　analyzed　and　explained　by　Explainable　Artificial　Intelligence　(XAI)　technology.　　©　1963-2012　IEEE.