In　pragmatic　engineering　milieus,　rotating　machinery　mostly　operates　under　normal　condition,　leading　to　the　long-tailed　monitoring　data　distribution　with　far　more　normal　than　fault　instances.　This　significant　class　imbalance　undermines　the　efficacy　of　standard　intelligent　fault　diagnosis　models.　Though　cost-sensitive　learning　helps,　two　challenges　remain:　1)　Existing　convolutional　neural　network　(CNN)　based　feature　extractors　struggle　to　capture　global　fault　information;　and　2)　current　cost-sensitive　losses　need　extensive　manual　tuning　of　sensitive　hyperparameters,　demanding　time　and　effort　while　being　user-unfriendly.　To　circumvent　such　issues,　a　novel　long-tailed　fault　diagnosis　framework　of　rotating　machinery　based　on　extended　attention　signal　transformer　with　adaptive　class　imbalance　loss　(EAST-ACIL)　is　proposed　in　this　paper.　The　lynchpin　innovations　are　threefold:　Primarily,　an　avant-garde　extended　attention　signal　transformer　(EAST)　is　constructed　to　extract　discriminative　representations　from　long-tailed　data.　In　EAST,　a　1-dimensional　(1D)　CNN　is　utilized　for　token　embedding　construction,　and　2D-CNN　for　developing　the　attention　extension　module,　thereby　mitigating　attention　smoothing　and　augmenting　the　model＇s　generalizability.　Secondly,　a　novel　adaptive　class　imbalance　loss　(ACIL)　is　designed　to　dynamically　reweight　training　data.　In　ACIL,　an　adaptive　class-level　weighting　term　automatically　accentuates　challenging-to-classify　categories　during　training,　while　a　boundary　regularization　term　maximizes　the　inter-class　margin,　substantially　increasing　the　model＇s　sensitivity　to　rare　fault　classes.　Lastly,　the　amalgamation　of　the　proposed　EAST　and　ACIL　modules　culminates　in　the　EAST-ACIL　diagnosis　framework.　Extensive　validation　on　rotor　and　bearing　fault　datasets　demonstrates　that　this　framework　surpasses　existing　methodologies　in　long-tail　fault　diagnosis,　achieving　superior　diagnosis　accuracy　even　under　extremely　imbalanced　conditions.