In　engineering,　labeled　data　acquired　from　bearings　is　extremely　scarce　due　to　its　high　cost.　Most　intelligent　diagnostic　methods　struggle　to　learn　sample　relationships　and　perform　poorly　without　sufficient　labeled　samples.　To　address　the　decline　in　accuracy　for　bearing　fault　diagnosis　with　limited　labeled　data　and　make　full　use　of　both　fault　information　from　unlabeled　data　and　class　information　from　labeled　data,　a　spatial–temporal　graph　attention　contrastive　learning　(STGACL)　is　developed　in　this　study.　In　the　STGACL,　a　triplet-channel　graph　construction　method　is　first　designed　to　generate　the　original　graph　for　semi-supervised　learning,　along　with　positive　and　negative　graphs　for　unsupervised　learning,　and　to　apply　label　propagation　for　the　utilization　of　class　information　from　labeled　data.　Second,　a　spatial–temporal　similar　graph　attention　network　is　constructed　as　the　backbone　network　to　initially　extract　graph　representations,　where　the　gated　convolution　captures　temporal　correlations,　and　the　similar　graph　attention　layer　is　employed　for　spatial　information　and　more　similar　nodes.　Finally,　graph　contrastive　learning　is　proposed　to　deeply　capture　additional　informative　features　from　unlabeled　samples,　and　a　new　loss　function　based　on　positive　and　negative　losses　fosters　the　model　to　fully　utilize　limited　labeled　samples.　The　STGACL　achieves　the　accuracy　of　98.07　%　and　93.88　%　on　the　aero-engine　bearing　and　planet　bearing　datasets　with　a　single　training　label,　surpassing　five　state-of-the-art　methods.　These　results　highlight　that　the　STGACL　excels　in　fault　diagnosis　for　limited　labeled　data,　with　promising　engineering　applications.　©　2025　Elsevier　Ltd