In　engineering,　imbalanced　data　collected　from　planet　bearings　causes　most　intelligent　models　to　shrink　the　decision　boundary　of　minor　classes　and　degrade　diagnostic　accuracy.　Different　from　these　models　under　the　assumption　of　data　balance,　graph-based　methods　focus　on　the　relationship　between　data　to　alleviate　the　issue　of　data　imbalance,　but　they　have　restrictions　on　single-feature　propagation　and　only　rely　on　the　feature　extraction　capability　of　convolutional　operations.　As　such,　a　multi-scale　dynamic　graph　mutual　information　network　(MDGMIN)　is　proposed　for　the　health　monitoring　of　planet　bearings　with　imbalanced　data.　First,　a　dual　spatial-temporal　graph　generation　algorithm　is　designed　to　construct　dynamic　and　distance　graphs　via　the　gated　convolution　in　the　temporal　dimension　and　the　cosine　similarity　and　Top-k　sorting　mechanism　in　the　spatial　dimension.　Second,　multi-scale　dynamic　edge　graph　convolutional　layers　are　constructed　to　extract　specific　and　similar　features,　and　they　are　weighted　fused　via　an　attention　mechanism.　Finally,　mutual　information　learning　is　developed　to　foster　the　model　in　capturing　graph　features　in-depth　through　commonality　and　discrepancy　constraints,　and　a　new　loss-driven　function　based　on　two　constraints　is　proposed　to　update　the　training　objective.　Experimental　analysis　on　an　imbalanced　planet　bearing　dataset　verifies　that　the　developed　MDGMIN　reaches　the　diagnostic　accuracy　of　92.80%,　exceeding　that　of　state-of-the-art　methods　on　the　dataset　with　an　imbalanced　ratio　of　20:1.　In　addition,　the　generalizability　of　the　MDGMIN　is　validated　in　another　bearing　dataset　from　the　planetary　gearbox.