With　the　rapid　development　of　artificial　intelligence　and　industrial　big　data　technologies,　unsupervised　transfer　learning　has　become　crucial　for　cross-domain　fault　diagnosis　of　rotating　equipment.　However,　most　existing　approaches　focus　on　global　domain　adaptation,　overlooking　the　fine-grained　distribution　discrepancy　among　subdomains.　For　this　reason,　this　paper　proposes　a　novel　Multi-Adversarial　Subdomain　Adaptation　Network　(MASAN)　for　fault　diagnosis　under　various　operating　conditions.　In　the　feature　extraction　module,　a　two-branch　feature　extractor　based　on　Graph　Convolutional　Network　(GCN)　paired　with　Improved　Bidirectional　Gated　Recurrent　Unit　(IBiGRU)　is　designed　to　integrates　spatial　information　from　graph-structured　data　and　temporal　features　from　sequential　data.　The　classifier　utilizes　the　extracted　domain-invariant　features　for　fault　type　identification.　In　the　domain　adaptation　module,　we　introduce　the　label　smoothing　strategy　into　the　Local　Maximum　Mean　Discrepancy　(LMMD)　to　mitigate　the　negative　impact　of　hard　labels　on　the　network＇s　generalization　ability.　Additionally,　multiple　domain　discriminators　are　integrated　to　optimally　align　the　fine-grained　distributions　of　the　source　and　target　domains.　Experimental　validation　on　the　Drive　Dynamic　Simulator　(DDS)　and　Paderborn　University　(PU)　dataset　further　highlights　its　exceptional　feature　extraction　and　domain　adaptation　capabilities.