In　the　actual　operation,　rotating　machinery　mostly　works　under　normal　condition.　The　collected　monitoring　data　often　exhibit　serious　distribution　imbalance　with　far　more　normal　label　samples　than　fault　label　samples,　leading　to　poor　recognition　performance　of　standard　intelligent　diagnosis　models.　Besides,　many　intelligent　diagnosis　models　rely　on　data　generation　to　overcome　this　problem,　which　is　subject　to　data　generation　differences.　Therefore,　to　address　above　limitations,　a　novel　temporal-spatial　multi-order　weighted　graph　convolution　network　(TSMOW-GCN)　with　refined　feature　topology　graph　is　proposed.　First,　a　multi-order　weight　graph　convolution　layer　is　proposed　to　aggregate　multi-order　weighted　mixing　neighbor　information　in　different　distances,　which　achieves　broader　representation　and　mines　more　features　and　relationships　without　data　generation　and　deep　network　structure.　Further,　the　feature　modeling　in　temporal　dimensions　is　considered.　Second,　a　refined　feature　topology　graph　construction　method　is　developed　to　obtain　compact　and　efficient　feature　topology　graphs,　which　can　improve　the　ability　of　graph　representation.　Besides,　a　dynamically　adjusted　label　smoothing　regularization　loss　is　proposed　to　further　improve　generalization　ability　and　avoid　overfitting　of　the　trained　model　under　imbalance　data.　Two　rotating　machinery　datasets　are　used　to　quantitatively　verify　proposed　method,　indicating　that　the　TSMOW-GCN　outperforms　several　advanced　approaches　under　various　imbalance　ratios.