In　deep　learning　model　based　on　multi-source　data　fusion,　features　of　different　types　of　signals　are　usually　mapped　to　fusion　layer　using　an　iso-specific　gravity　method.　However,　this　process　ignores　the　problem　of　inconsistent　contributions　of　non-homologous　signal　features　to　the　final　recognition　effect.　Here,　a　deep　learning　model　based　on　dual-attention　mechanism　was　proposed.　Firstly,　this　model　could　use　a　channel　attention　module　to　suppress　effects　of　irrelevant　components　in　homologous　signal,　and　then　use　a　multi-source　data　attention　module　to　adaptively　assign　weights　to　non-homologous　signal　features.　Then,　the　re-weighted　features　were　fused.　Finally,　a　classifier　was　used　to　realize　pattern　classification.　The　proposed　method　was　applied　in　fault　diagnosis　of　induction　motor.　The　results　showed　that　the　average　recognition　accuracy　of　this　method　is　99.　74%,　its　diagnostic　effect　is　better　than　existing　methods＇.　©　2023　Chinese　Vibration　Engineering　Society.　All　rights　reserved.