Unsupervised　Domain　Adaptation　(UDA)　has　been　widely　used　for　fault　diagnosis　to　solve　data　distribution　shifts　in　multi-source　domains.　However,　UDA　ignores　the　imbalance　of　the　class　distribution　in　the　domain,　which　makes　it　difficult　to　apply　to　real　industrial　processes.　Therefore,　we　propose　a　class-aware　multi-source　domain　adaptation　(CMDA)　for　imbalanced　fault　diagnosis.　Firstly,　a　batch　interaction　feature　extractor　is　designed　to　apply　the　cross-attention　mechanism　to　the　batch　dimension　to　capture　nonlinear　relationships　among　imbalanced　features　of　classes.　Secondly,　we　construct　a　class-aware　UDA　module　that　is　introduced　with　class　information　in　both　the　feature　and　discriminant　layers　to　realize　domain　adaptation.　The　module　performs　multilinear　augmentation　of　imbalanced　class　samples　for　data　alignment.　Finally,　a　task　classifier　is　designed　that　combines　multiple　classifiers　with　reliability-weighted　scores　to　weaken　negative　transfer　and　jointly　accomplish　fault　diagnosis　of　rolling　bearings.　The　reliability-weighted　scores　are　determined　by　the　inter-domain　distance　and　diagnostic　accuracy.　We　conducted　experiments　using　the　CWRU　dataset　under　4　protocols　and　12　scenarios　and　verified　the　effectiveness　of　CMDA.